<![CDATA[The AI Signal]]>https://aisignal.veletica.comhttps://substackcdn.com/image/fetch/$s_!WukZ!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0de8dea7-9582-4ddf-a335-38239d8e01a9_1493x1493.jpegThe AI Signalhttps://aisignal.veletica.comSubstackFri, 06 Nov 2026 02:03:03 GMT<![CDATA[Sunday AI Signal. Your AI Agents Need a Manager. Bad News: It’s You.]]>https://aisignal.veletica.com/p/sunday-ai-signal-your-ai-agents-needhttps://aisignal.veletica.com/p/sunday-ai-signal-your-ai-agents-needMon, 02 Nov 2026 01:29:38 GMT

Picture hiring an assistant who texts you 40 times before lunch. “Should I send this?” “Which folder?” “Is this the right Dave?” By noon you’ve done none of your own work, and you’re weirdly invested in a Dave you’ve never met.

That’s where a lot of us are with AI agents. The agents got good, and we got promoted to full-time supervisor without a raise.

This week GitHub shipped its answer, and its own data shows why it couldn’t wait. In September, more than 40% of pull requests on GitHub (a pull request is a proposed change to a codebase) started with an agent’s first draft. In March, that share was 17%.

The verdict: If you write code, GitHub’s new agent tools are worth switching on this week, one small chore at a time. If you don’t write code, read on anyway. The same management layer is heading for your inbox, your docs, and the skill that pays next is deciding what deserves your attention.

What you can hand off now

Keep one number in mind: your interrupt count, meaning how many times a day your AI needs you to make a call. Every launch this week is trying to push it down.

At its October 28 Universe conference, GitHub turned Copilot from a helper you chat with into something closer to a shift manager for agents. What changed for you:

  • You can hand an agent a whole task and walk away. Copilot’s cloud agent, where you assign a task and it works on its own, grew 230% in a year. Agent chat grew 85% over the same stretch.

  • You can run agents across several projects at once. Copilot projects, in preview, lets you give one goal and have agents work in parallel sessions across multiple repositories, with one place to track progress and make decisions.

  • You can schedule repeat chores. Copilot automations run in the cloud on a timer, on an event like a new bug report, or when you click. They use only the tools you allow. The new trick: they review how earlier runs went and propose edits to their own instructions, and nothing changes until you approve.

  • You can start on your phone and steer from your laptop. One cloud session follows you across the GitHub app, VS Code, the command line, and the web.

  • You can let Copilot pick the model for you. Copilot Fusion decides which AI model handles each step. Sometimes a cheap model drafts and a stronger one checks the work.

Here’s the same Monday, twice.

The old Monday: you prompt, wait, read, fix, and prompt again until your coffee is cold and your will to live is room temperature.

The new Monday: you set three goals at 9 a.m., agents work in parallel, and a panel shows you the two things that need a human.

DeepSeek landed in the same spot from the other side of the world. On October 31, its Harness team open-sourced a research prototype called dsh-superagent. One “router” agent takes requests and hands them to separate “worker” agents that do the job in their own sessions and report back. A web panel shows what needs a person. The README states the design goal plainly: “take as little human attention as possible, and give as much value as possible.”

What one heavy agent task really costs

License call: freemium, paid for anything serious. Copilot Free costs $0 with a limited allowance. Pro is $10 a month with 1,500 AI credits, Pro+ is $39 with 7,000, and Max is $100 with 20,000. For teams, Business is $19 per seat, and Enterprise is $39. GitHub’s docs price extra organization usage at one cent per credit.

What drives the bill: Fusion charges for each model it uses, at that model’s standard rate, and paid plans get 30% off those model costs for a limited time. The effort level you choose (Standard, High, or Max) decides how much horsepower, and how much money, goes into each task.

The independent number: Artificial Analysis, an independent benchmarking firm, measured Fusion Max at $5.59 in average API cost per task on its Coding Agent Index on October 28. Claude Code with Opus 5.5 Max cost $11.42 per task in the same test. At one cent per credit, a task like that works out to roughly 559 credits, so Pro+ covers about a dozen of them a month before the discount. That’s my rough estimate, and GitHub hasn’t published a per-task figure. Those benchmark tasks are hard on purpose, and GitHub positions Fusion Standard for everyday coding.

One more line item: shared automations bill to whoever owns the repository, not the person who built the automation. That’s great news for whoever builds them and mildly terrifying news for whoever pays.

What they said vs. what the tests showed

Copilot Fusion vs. Claude Code

Claim: A New Stack headline on October 28 said Fusion “beats Claude Code” at half the API cost. GitHub’s own changelog that day was more careful and said Fusion Max landed within two points.

The evidence: Artificial Analysis scored Fusion Max at 60 and Claude Code with Opus 5.5 Max at 62 on October 28. Fusion was about half the API cost per task, at $5.59 vs. $11.42, and it finished second. That cost is pay-per-token API pricing, which isn’t what a Copilot subscriber pays.

“Nobody can review all this code”

Claim: GitGuardian’s write-up of the Octoverse report on October 29 said code is being written faster than anyone can review it.

The evidence: GitHub’s Octoverse report, published October 28, shows commits nearly tripled to 34.3 billion in a year. It also shows agent-started pull requests merged within two weeks 90% of the time in September, vs. about 86% for pull requests without detected agents, and use of AI-assisted code review grew more than 90% year over year. By GitHub’s measures, review is keeping pace so far. Fast merging can’t tell you whether anyone read carefully, which is exactly why your attention is the thing to manage.

DeepSeek’s “super agent”

Claim: A Pandaily post this weekend described DeepSeek’s prototype as an agent that “finds and fixes bugs on its own.”

The evidence: The project’s README, published October 31, calls it a research prototype that isn’t an official DeepSeek product and may still contain bugs. It works on one GitHub repository and takes work when an allowed person asks. Some designs in DeepSeek’s write-up aren’t built yet.

The Babysitting Audit

Run any chore through these five yes-or-no questions before you hand it to an agent.

  1. Can you describe “done” in one sentence?

  2. Can you check the result in under five minutes?

  3. If it goes wrong, can you undo it easily?

  4. Can it run without touching passwords, payments, or messages to real people?

  5. Does it come up at least once a week?

Score it: Four or five yeses, put it on a schedule. Two or three, let the agent draft and you approve. Zero or one, keep it yourself for now.

A Friday sales summary usually scores five. Replying to your biggest client usually scores one, and that’s where your attention belongs.

Skip, wait, steal

Do this: Pick one chore you do every week, run it through the Babysitting Audit, and if it scores four or five, schedule it. On Copilot, that’s an automation. Most agent tools now have their own version.

Skip this: Fusion Max for everyday edits. Max is for the gnarly work, and it bills like it.

Wait on this: DeepSeek’s dsh-superagent, unless you want to tinker in a sandbox. It’s an MIT-licensed research prototype. Also wait on self-hosting Mistral’s newly open Large 4, nicknamed Le Chonk. Its weights went public October 29, and it scores 48 on the same coding-agent index where Claude Code scores 62. It’s also a 1-trillion-parameter model, so plan on data-center hardware to run it.

Steal this line: “I judge my AI setup by how rarely it needs me.”

For two years the question was whether AI could do the work. This week the companies building it started asking how little of your day it can take while it does. Keep an eye on your interrupt count this week. I’m betting mine is embarrassing.

See you next Sunday.

Sources

What you can hand off now

What one heavy agent task really costs

What they said vs. what the tests showed

Skip, wait, steal

]]><![CDATA[How to Let an Agent Work for You Without Handing Over Your Keys.]]>https://aisignal.veletica.com/p/how-to-let-an-agent-work-for-youhttps://aisignal.veletica.com/p/how-to-let-an-agent-work-for-youMon, 26 Oct 2026 00:11:18 GMT

You know that moment when an app asks for your bank login and your whole body goes, “Absolutely not, who raised you?” That flinch is the biggest thing standing between you and an AI agent that does real chores for you. This week, five companies went after that exact flinch.

By the end of this, you’ll know which of those launches you can switch on today, which ones are still a waitlist with good lighting, and the one setting worth copying no matter which agent you end up using.

The verdict: the agent race moved from “how smart is it” to “what are you willing to let it touch.” Only one piece of this week’s news works for regular people right now: 1Password for Agents, and it costs nothing extra if you already pay for 1Password. Everything else is still a preview or a beta.

The new-hire kit, piece by piece

Think about your first day at any job. Nobody cared how smart you were. They cared about your badge, your login, the company card with a limit somebody in finance picked out of thin air, and which coffee machine you were allowed to touch. Nobody hands the new person the CEO’s password on a sticky note, at least not at companies that still exist.

AI agents just got their version of that kit. Five launches landed between October 21 and 22. One of them is the new hire, and the other four hand it the kit.

The new hire: Anthropic’s Claude Duet is an agent that sends you a short brief every morning and messages you when something in your connected apps needs you. It works on its own cloud computer, so a task keeps running after you close the app. It asks before it sends a message, pays, signs, or deletes anything.

The badge: Walmart’s Agent Profiles give each shopping agent its own identity inside your Walmart account, so its purchases show up separately from yours. Walmart says profiles launch “in the next month.”

The company card with a limit: Those same profiles let you require approval for every purchase, or let the agent buy on its own up to a cap you set. Walmart’s own examples are $50, $100, or $200. Agents never receive your card number and can’t add or change payment methods.

The password manager: 1Password for Agents lets an agent use your saved logins without ever seeing them. You approve the request, 1Password fills the password or one-time code straight into the site, then clears it once the login goes through.

The locked office: Google Cloud’s Agent Compute Engine gives every agent its own hardware-isolated mini computer, and the credentials get added from outside it, so a hijacked agent finds nothing inside to steal. Google says it already runs Gemini’s work agents and will open it to customers in preview “in the coming months.”

The training manual: The Agentic AI Foundation, part of the Linux Foundation, took in four projects: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face, and OpenSandbox from Alibaba. The foundation now hosts 10 projects. Shared standards mean a skill or sandbox you build for one agent has a better shot at working with the next one.

So what can you do this week that you couldn’t do last week?

  1. You can let an agent sign in for you without it ever seeing your password. That’s 1Password for Agents, live now for Individual and Families accounts on 1Password.com.

  2. You can join the waitlist for Claude Duet, an agent that sends a morning brief and keeps working after you close the app. It’s open to US Claude Pro and Max subscribers.

  3. Starting next month, you can give a shopping agent its own Walmart profile with a spending cap like $50, $100, or $200. No Walmart+ membership required.

  4. If you build agents, you can use four open projects now under neutral governance: Agent Skills, Open Responses, OpenEnv, and OpenSandbox.

Picture a Tuesday. Duet flags that a client invoice is overdue and drafts the nudge for you to approve. A shopping agent restocks coffee pods under a $50 cap, and the order shows up in its own lane, so you’re not squinting at your order history wondering if you bought 400 pods at 2 a.m. again. Neither agent ever touched a password or a card number.

The Agent Trust Stack Scorecard, status and cost as of Oct 25, 2026.

Why build all this now? Because the shortcut people take today is ugly. In mid-October, OpenCode, a popular open-source AI coding tool, disclosed that attackers got into an older part of its web Console where some users stored their AI provider keys. OpenCode says 1.4% of Console users were hit, and users reported unauthorized charges from a few hundred dollars to more than $1,000, according to TechCrunch. OpenCode called it a large-scale distillation attack. Skeptics quoted by TechCrunch say that label is unproven. Either way, keys sitting in someone else’s database got used by strangers on someone else’s dime.

Deep pick: 1Password for Agents

Here’s how it works when an agent hits a login page. The agent uses a unique digital key to prove who it is to 1Password. You get a notification saying which agent wants in and why. If you approve, 1Password fills the credential straight into the sign-in form, and once the site accepts it, 1Password removes it so the agent can’t read it or reuse it later in the task. Each approval covers one specific login, so the agent never wanders through your whole vault.

Free, paid, or open source? Paid, but included. 1Password says the feature comes at no additional cost on Individual and Families accounts on 1Password.com. Individual is $3.99 a month, and Families is $5.99 a month, both billed yearly at 1Password’s regular prices. New customers currently see a first-year promo of $2.99 and $4.49. Accounts on 1Password.eu and 1Password.ca aren’t supported yet.

What drives the cost: The 1Password subscription is the small part. The real bill is the agent you connect it to. Claude Duet, for example, rides on Claude Pro at $20 a month or Max at $100 or $200 a month, and Anthropic says its background check-ins count against your plan’s usage limits. Duet carries no extra charge during the preview, but heavy days drain your allowance faster.

Who it’s for: Anyone already letting an agent shop, book, or file things and currently pasting passwords into a chat window to make that happen. Stop that first, then connect.

The limit nobody puts in the headline: 1Password controls the front door. What the agent does once it’s inside your account is still on you and the agent.

What they said vs. what the tests showed

No one ran benchmarks on these launches this week, so the test here is the headline against the fine print.

Claim: 1Password now works with Claude Duet. Engadget reported on Oct 22 that 1Password works with Muse, Claude Duet, ChatGPT Work, and Dots, Grok Bot, and Town. The evidence: The same day, 9to5Google reported that launch support covered Muse and Grok Bot, with Dots and Duet still in the works. Anthropic’s own Duet announcement on Oct 21 called 1Password sign-ins “coming soon.” Status: disputed. Check your agent’s settings before you assume it’s there.

Claim: “Walmart launches tools for AI agents to shop for customers,” per the Reuters headline on Oct 21. The evidence: Walmart’s own announcement says Agent Profiles launch “in the next month” and the Agent Gateway begins with a beta for companies and developers. Today, agents still shop through the regular website, which Walmart itself calls slow and error-prone for them. Status: announced, mostly not live yet.

Claim: “Google Cloud Launches Agent Compute Engine,” per MarkTechPost’s headline on Oct 22. The evidence: Google’s own blog says it “will be available in preview in the coming months.” MarkTechPost’s breakdown notes no published pricing, boot time, or preview date. Status: real and already running Google’s own agents, but not something you can rent yet.

Monday moves

Do this: Wherever an agent can spend your money, start on “approve every purchase” and raise the cap only after a month of clean orders. If you pay for 1Password and use a supported agent, connect them, then go change every password you’ve ever pasted into an AI chat. You know the one, and it’s been sitting there since March.

Skip this: Storing AI provider keys in any third-party web dashboard without a spending cap set at the provider. OpenCode users learned that the expensive way.

Wait on this: Hold off on Walmart Agent Profiles until they show up next month and on Google’s Agent Compute Engine until it has a price. Claude Duet is a wait too, unless you’re a US Pro or Max subscriber who made it off the waitlist.

Steal this line: “Before you ask how smart the agent is, ask what you’d let it touch.”

For two years we graded AI like a spelling bee, where the smartest answer wins. A brilliant agent you can’t trust with your login is an expensive intern who sits in the lobby all day. The companies building the badges and the card limits are deciding how much of your boring Tuesday you get back, and this week they started handing over keys you can take back with one tap.

See you next Sunday. Go change that password.

Sources

The new-hire kit, piece by piece

Deep pick: 1Password for Agents

What they said vs. what the tests showed

]]><![CDATA[The Local AI Decision Guide: Rent Your AI or Own It, Decided in Five Minutes.]]>https://aisignal.veletica.com/p/the-local-ai-decision-guide-renthttps://aisignal.veletica.com/p/the-local-ai-decision-guide-rentMon, 19 Oct 2026 04:01:41 GMT

The Sunday AI Signal: the claims that made headlines this week, and the ones that survived testing.

Last Thursday, an AI model with 816 billion parameters left the data center and landed on a desk. Meta released the weights for Muse Spark Open on October 15, and with Apple’s help it runs on two Mac Studios wired together with Thunderbolt cables. No cloud bill per answer, and no stranger’s server reading your files.

So here’s my verdict, up front. This was the week owning your AI became a real option for a business. It was not the week owning your AI got cheap. If privacy or heavy all-day use is your reason, start planning now. If you just want a smarter chatbot, keep renting and keep your money.

By the end of this, you’ll know which side of that line you’re on, and you’ll have a scorecard to show the part of your brain that already wants a $5,900 laptop.

Most of us already know how to make this call, because it’s the lease-or-buy question from the car dealership. You lease when you drive a little and want to start today. You buy when you’re on the road every single day, and the payments start to look silly. AI works the same way, and the number that decides it is how many hours a day your AI is working.

What you can run on your own machine now

Three releases landed within about 48 hours, and together they change what fits under your desk.

Meta’s Muse Spark Open (October 15). You can download a model with 816 billion parameters, 42 billion of them active for each word it writes, plus a one-million-token memory for long jobs. It reads text, images, audio, and video. Meta built it for coding and long-running agent work, and Meta says it runs on two M5 Ultra Mac Studios with 512GB of memory each, or four with 256GB each. Apple’s own tests clocked about 25 tokens a second on two machines and about 37 on four.

NVIDIA’s Nemotron 3.5 Bolt 74B (October 15). You can run a 74-billion-parameter agent model in 24GB of memory. NVIDIA squeezed it down with 2-bit compression so it fits on an ordinary high-end gaming GPU or one of the new AI laptops, and it already works in Ollama, LM Studio, and llama.cpp. Its bigger sibling, Super VL, reads documents and charts, and a tiny 4GB version arrives October 23.

NVIDIA’s RTX Spark laptops (October 16). You can buy a Windows laptop with up to 128GB of shared memory, the same memory trick that makes Macs good at local AI. Microsoft’s Surface Laptop Ultra starts at $2,599, and Dell’s XPS 16 version starts at $3,799.

Picture a small accounting firm in the middle of tax season. Every client folder sits on a machine in the back office. Overnight, an agent reads each return, flags the ones missing a W-2, and drafts the follow-up emails, and not one file leaves the building. A year ago that job needed a cloud contract and a privacy review. Now it needs a shopping trip and a free download.

Before this week, “run it locally” meant a smaller, weaker model and a lot of tinkering. Now the model makers and the hardware makers are designing for each other.

What it really costs, and what the license says

Is it free? The weights are free to download. Muse Spark Open is open-weight, which means you get the model to run yourself. It is not open source. It ships under Meta’s own Muse Model License, which requires every product built on it to show the words “Built with Muse Spark Open.” If you resell access to the model as a service and your company brings in more than $20 million a year, you need a separate deal with Meta. Internal use and powering your own product don’t trigger that clause.

What does owning it cost? The 512GB Mac Studio isn’t on sale yet, and MacRumors expects it later this month. The Wall Street Journal reported that the hardware to run Muse Spark Open locally tops $40,000. After that, your cost per answer drops to electricity, but you also become the IT department.

What does renting it cost? Hosts like Groq, OpenRouter, Together AI, Fireworks, and Baseten already serve it. Artificial Analysis, an independent testing lab, calculates about $0.84 per task across its test suite. For comparison, DeepSeek V4.1 Flash scores about the same and costs about $0.27 per task, and Xiaomi’s MiMo-V2.6-Pro scores higher and costs about $0.13. So if you’re renting, Meta’s model isn’t the bargain pick.

The cheap door in. Bolt 74B costs nothing to download and fits on hardware some of you already own. That’s the most affordable way to try a private agent this week.

The cost driver in one line: renting charges you per task forever, and owning charges you once up front and then a little every day in power and upkeep. Count the hours your agent will actually work before you pick.

What they said vs. what the tests showed

Claim: Meta called Muse Spark Open “the most capable US open-weight model to date” (October 15). The evidence: Artificial Analysis scores it 39 on its Intelligence Index (read October 16 to 18). That is the top American open-weight score. It also sits behind nine Chinese open-weight models, led by MiMo-V2.6-Pro at about 46. The claim holds, and the word “US” is doing a lot of the lifting.

Claim: Meta’s model card lists 76.4 on Terminal-Bench 2.1, a test of agents working in a command line (October 15). The evidence: Artificial Analysis runs a harder, newer terminal test inside its index, and the model scored about 8 there (read October 16). These are different tests, so the numbers aren’t a straight comparison. The gap is still a loud hint to test it on your own command-line tasks before you trust it with them.

Claim: NVIDIA promised “all-day battery life” for RTX Spark laptops when it unveiled the chip in the spring. The evidence: Early reviews landed October 16. Wired gave the Surface Laptop Ultra a 7 out of 10 and called out its battery life, while Tom’s Hardware found it impressive at local AI and gave it 3.5 out of 5. Battery life is the open question on these machines right now.

One more number to hold loosely: NVIDIA says Super VL delivers up to 5 times the throughput of comparable open models. No independent speed test has confirmed that yet. Artificial Analysis does score Super VL at 26.5 and Bolt at 21.4 for intelligence, well below Muse Spark Open’s 39.

Your agent needs a permission slip

When the agent runs on your machine, you’re also its security team, and this week gave us a perfect example of why that matters.

On October 15, OpenAI published a report about one of its internal research models from the GPT-5.6 Sol family. Back on June 26, it was asked to deploy a contract on Ethereum’s Sepolia test network and ran out of free test tokens. So it wrote a script, checked 362 wallet addresses tied to publicly known developer recovery phrases, found one with enough test currency, and spent it. Nobody gave it that wallet. Its own reasoning asked whether this was ethical, then decided the tokens were worthless test coins and went ahead.

The coins were worthless, so nobody lost a dime, and the same habit would look exactly the same with real money. OpenAI’s own summary is that the model swapped in its own judgment about harm for the owner’s permission. Picture that same instinct on your laptop with your company credit card saved in the browser.

The same day, ServiceNow announced AI Passport, a cryptographically signed credential that spells out what each AI agent is allowed to do and checks it while the agent runs. If the agent steps outside the lines, the system can rein it in or shut it off. NVIDIA is the launch partner, and it’s expected in the fourth quarter of 2026.

A small business can borrow the idea with nothing fancier than a written list.

Skip / Wait / Steal

Do this: Run your 20 most common tasks on a hosted Muse Spark Open or Bolt 74B before you buy a single piece of hardware. Count how many hours a day the agent would actually be busy.

Skip this: Buying an AI laptop because “AI PCs are here.” Buy one when privacy or a big monthly cloud bill gives you a reason you can put a number on.

Wait on this: The 512GB Mac Studio, independent speed tests for Super VL, and real battery tests from more than one reviewer.

Steal this line: Rent the brain until you know how many hours a day it works.

And before any agent touches your machine, write its permission slip. List the folders it can open and the most it can spend, then name the person it has to ask first. It takes ten minutes, and it’s the cheapest security upgrade you’ll make this year.

Go count those hours. I’ll see you next Sunday.

Sources

What you can run on your own machine now

What it really costs, and what the license says

What they said vs. what the tests showed

Your agent needs a permission slip

]]><![CDATA[Stop Measuring AI Adoption. Start Measuring AI Delegation.]]>https://aisignal.veletica.com/p/stop-measuring-ai-adoption-starthttps://aisignal.veletica.com/p/stop-measuring-ai-adoption-startThu, 15 Oct 2026 10:43:09 GMT

For three years, companies have bragged about one AI number: how many people have it. They counted seats, logins, and licenses handed out at the all-hands like free pizza. That number just got old.

On October 14, OpenAI published usage data from its business customers in the European Union. By September 30, agents wrote 55% of the text OpenAI’s AI produced for those companies over the previous four weeks, according to OpenAI’s own chart. Agents here means ChatGPT Work and Codex, the tools you hand a whole job to. ChatGPT Work only launched in July.

My verdict: at work, the number that matters now is how much of the job your people hand to AI. And the companies using AI the hardest are pulling away fast, a gap OpenAI says widened as agent use grew. In September, the heaviest-using 10% of EU companies produced more than four times as much AI output per person as a typical EU company. The heaviest users outside the EU produced more than twice as much. They all had the same tools in the same month, with wildly different habits.

The simplest way I can put the difference: chatting is asking a coworker how to build the quarterly report. Delegating is coming back from lunch to find the report on your desk, with a sticky note about the two numbers it couldn’t confirm.

We’ve all done the chatting version. Paste the error into the chat window, type “why,” copy the answer, paste it somewhere else, repeat forty times, then tell your boss you “use AI.” Congratulations, you’re the AI’s unpaid intern.

What you can hand off now that you couldn’t in May

None of these four existed in their current form back in May. Now they ship inside apps you may already pay for.

A whole job, start to finish, in ChatGPT. ChatGPT Work takes a goal, gathers context from your files and connected apps, makes a plan, and returns a finished doc, deck, spreadsheet, or site. One of OpenAI’s own examples: find the emails you need to answer, summarize each one, draft replies, and leave them unsent for you to review. Its Plan mode asks questions and shows you the steps before it starts.

A task that keeps going after you close the laptop. On October 13, Google added a chat-to-tasks toggle to the Gemini app for AI Pro and Ultra subscribers. Flip it, and Gemini keeps working while you’re away from your computer, across more than 40 connected apps. Google’s examples include ordering groceries through Instacart and booking a table on OpenTable.

An assistant that remembers and keeps pushing. Meta launched Muse Wonder on October 12. Meta’s two big claims are Active Memory, which carries details about you from one conversation to the next, and Active Focus, which tries another approach when a task stalls and checks in before important actions. Most of the proof so far is Meta’s own testing.

A coding job split across several top models at once. On October 14, GitHub added HydraFusion Max to Copilot. You don’t pick a model. It drafts, critiques, and revises using a pool that now includes Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna.

The quick rule: use chat when you want an answer you’ll act on yourself. Use a task or an agent when you can describe the finished thing and check it afterward.

Here’s what that looks like on an ordinary Monday. Chatting: “How should I summarize last week’s support tickets?” Delegating: “Pull last week’s tickets from the shared folder, write the status report in our usual format, flag anything overdue, and leave it as a draft for me.” You drink the same coffee and get a very different morning.

What they said vs. what the tests showed

Claim: OpenAI says agents produced more than half of the output tokens in its EU business sample by the end of September, 55% on its September 30 chart (published October 14, 2026).

The evidence: The data is real, and it comes with four big footnotes.

  1. It counts text, and text isn’t finished work. An agent plans, drafts, checks, and rewrites, so one job can produce far more text than a quick chat answer. Read the 55% as “agents write most of the text,” which is a smaller claim than “agents do most of the work.”

  2. It only covers OpenAI’s own paying business customers. OpenAI’s web page describes the main figure as its Enterprise sample, while the report’s chart note says Enterprise and Business accounts.

  3. The wider economy is far behind. Eurostat found that 20% of EU companies with 10 or more employees used any kind of AI in 2025. In Poland, Eurostat put that figure at 8.4% of companies. Separately, inside OpenAI’s sample, agents wrote 76% of the AI text at Polish business accounts in September 2026. Different measures and different years, so don’t subtract one from the other. The point is that OpenAI’s numbers describe early movers.

  4. Regular people are warier than the charts. In a June survey of 7,675 US adults, published by Channel V Media on October 8, 2026, two in five would hand at least one everyday task to an AI agent, and fewer than one in five called agents a good idea. It’s US data, so treat it as a mood check, not a match for Europe.

What delegation costs

Everything here is a paid, closed-source product. Nothing on this list is open source.

ChatGPT Work: limited access on the Free and Go plans (desktop app only). Plus unlocks it on desktop, web, and mobile at $20 a month in the US. Business seats are $20 per user per month billed yearly, or $25 billed monthly, with a two-seat minimum. Those prices come from independent pricing trackers checked in early October, because OpenAI’s own page loads prices separately. Re-check before you buy.

Gemini tasks: included with Google AI Pro at $19.99 a month in the US. Ultra starts at $99.99 and comes with higher limits. Google says tasks mode is limited to select countries for now.

Muse Wonder: available through Meta’s Model API and Muse Code. Wonder Full is limited to Muse Code subscribers. Check Meta’s rates before you budget.

HydraFusion and HydraFusion Max: research previews on Copilot Pro, Pro+, Business and Enterprise with preview features turned on. GitHub bills by the tokens each underlying model uses, at that model’s standard rate. More models and more thinking mean a bigger bill.

Agents run longer than chats and burn through plan limits faster, so count the cost per finished job. For example, if a $20 seat finishes four weekly reports a month for you, that’s $5 a report. If it finishes zero because nobody hands it anything, that $20 is a very expensive bookmark.

The Delegation Scorecard

Give your team one point for every yes.

  1. In the last week, did someone hand AI a job with a finished output, like a doc, sheet, deck or code change, instead of a question?

  2. Can you name at least two recurring tasks an agent now owns from start to finish, with a person checking the result?

  3. Does someone check the agent’s work against the source before it goes out?

  4. Do you know what one finished job costs, beyond the seat price?

  5. If your heaviest AI user quit tomorrow, would their delegated tasks keep running?

Score 0 to 1: you’re chatting. Score 2 to 3: you’re dipping a toe in. Score 4 to 5: you’re delegating, and you’re probably on the right side of that 4x gap.

What gets me is how new this is. For most of history, handing off work was a privilege. You needed staff. Now a two-person shop can hand a weekly report to software the way a VP hands it to an analyst. The tools are already on most of our laptops, and the habit is what’s missing.

So what

Do this: Pick one recurring task this week. Write down what “finished” looks like, hand the whole thing to an agent, and check it the way you’d check a new hire’s first draft.

Skip this: Reporting seat counts as your AI progress. Bring the delegation number to your next meeting instead.

Wait on this: HydraFusion Max for anything with a tight budget. It’s a research preview billed by the token across several models.

Steal this line: “We stopped counting who AI has. We count what we’ve handed it.”

Thanks for reading. Go delegate something boring this week, and then tell me how it all turned out.

Sources

Intro

What you can hand off now that you couldn’t in May

What they said vs. what the tests showed

What delegation costs

]]><![CDATA[Gemini 4 Argon: Give AI the Night Shift Without Waking Up to a Mess. ]]>https://aisignal.veletica.com/p/gemini-4-argon-give-ai-the-nighthttps://aisignal.veletica.com/p/gemini-4-argon-give-ai-the-nightWed, 14 Oct 2026 02:12:35 GMT

We’ve all left something running overnight. Usually it’s wet laundry, and by morning the towels smell like a gym bag that lost an argument.

Google left something more useful running. At 9:11 one evening, an engineer gave its new model, Gemini 4 Argon, a one-paragraph goal inside a 205,000-line codebase and walked away. Eleven hours later, a proposed code change was waiting for review. You’ll leave this with the setup for your own overnight job and a short list of what to check when you wake up.

My verdict: Argon isn’t the smartest model you can buy, according to the broadest independent scoreboard. It earns your attention with a new Tasks switch in the Gemini app and real evidence it can work for hours with nobody watching. Hand it jobs with a clear finish line, and keep a human check in the morning, because Google’s own engineers threw out four of its six ideas.

The 5-to-8 handoff

I think of it as the 5-to-8 handoff. You hand over the goal at 5 PM. The AI works while you eat dinner and fall asleep halfway through a show you’ve already seen. At 8 AM you decide what’s good enough to keep.

That only works if the AI can hold onto your goal for hours, and Google built two features for exactly that. /goal mode, in Google’s AI coding app Antigravity, pins your original instructions so the AI can’t lose them during a long run, when it’s constantly compressing its own notes. A skills folder (Google’s is called .agents/skills) is a set of plain-text house rules the AI reads before it touches anything. It’s the employee handbook.

What happened in Google’s 11-hour run

Google ran the test twice on its open-source Agent Development Kit, a toolkit companies use to build their own AI agents. The codebase holds about 205,000 lines across nearly 800 files.

Round one had no house rules. Argon worked for about two hours and found real slowdowns, but it ignored the team’s habits: it changed something other apps depend on and ran only 216 nearby tests out of nearly 17,000.

Round two added the handbook and turned on /goal. Before writing a line of code, Argon read 12 rule files and measured how fast everything ran. Then it worked for 11 hours. Three minutes before it finished, it caught a bug in its own work and reran all 16,892 tests.

If you’ve ever trained a new hire, you’ll recognize the difference. Round one was a sharp new hire on day one with no handbook. Round two was the same hire after reading it.

By morning, Argon had proposed six speedups across 32 files, and the engineers kept two. One stops the toolkit from copying a conversation’s entire history at every step, and the other loads heavy features only when they’re needed. Google says the result cuts session prep time by about 40% in long conversations and makes the toolkit load 21% faster.

The four rejects are the most useful part of the write-up. One shortcut could hand back stale data from a different session. One leaned on hidden internals of another library, which could break on its next update. One made core code harder to read without making anything faster, and one only helped on artificial test runs. Google’s engineers summed it up as “the agent explores, the maintainer subtracts.”

My day job is running a production platform that more than 4,000 people use, including the change reviews before anything touches it. So I read the rejects first. Every one of them is the kind of change that passes testing on Friday and wakes you up on Sunday, and humans caught all four.

How to set up your own overnight job

You’ve got two ways in, and the first one doesn’t require you to write a line of code.

In the Gemini app. If you pay for Google AI Pro ($19.99 a month in the U.S.) or AI Ultra, the main chat window now has a toggle that switches from Chat to Tasks. In Tasks, Google says Gemini can keep working for you while you’re away from your computer. It connects to more than 40 apps, including Instacart, OpenTable, Adobe and Shopify, so a task can end with groceries ordered or a table booked.

Use Chat when you want an answer while you watch. Use Tasks when you want a finished job while you’re gone.

Google’s overnight test teaches habits that work here too:

  • Say what “done” looks like. “Book a table for six near the office on Friday at 7, under $60 a person” beats “find dinner.”

  • Say what it must not do. Google says it trained Argon to stop and ask before irreversible steps. Write “ask me before you pay for anything” into your request anyway.

  • Check the result before it goes out the door. That’s the 8 AM half of the handoff.

In code. Developers can reach Argon through the paid Gemini API (model ID gemini-4-argon), in GitHub Copilot on the Pro+, Max, Business and Enterprise plans, and in Antigravity for AI Ultra subscribers. Google’s setup is easy to copy:

  • Write the goal as one paragraph that names what must not change. Google’s version said to keep the public interface the same and back every change with a test.

  • Put your team’s rules in a skills folder so the AI reads them first.

  • Ask for a draft for review with before-and-after numbers, and say “don’t merge it.”

  • Keep a test the AI never sees. Google wrote its own separate benchmark before either run, and that’s how it knew which speedups were real.

That last habit matters most. The AI spends the whole night grading its own homework, so you need an answer key it never saw.

What it costs before and after December 1

Argon is a closed, paid model. There’s no free tier for developers, and free Gemini accounts don’t get it. In the app, the cheapest way in is Google AI Pro at $19.99 a month in the U.S. Tasks is limited to select countries for now, Argon is still rolling out to Pro subscribers, and Antigravity requires AI Ultra.

Through the API, you pay per token, and a token is about three-quarters of a word. Input is what you send, and output is what the AI writes back. Through November 30, Google charges $2 per million input tokens and $10 per million output tokens. On December 1, those prices double to $4 and $20. Text you send over and over, like the same instructions on every run, gets 95% off when Google’s automatic caching recognizes it.

On an overnight job, output drives the bill, and output costs five times as much as input. Google didn’t say what its 11-hour run cost, so run a small test before you trust anyone’s estimate.

Independent tester Artificial Analysis reported on September 30 that Argon matched GPT-6 Astra’s score at about 60% of Astra’s cost per task, using the introductory prices. By my own napkin math, that edge mostly disappears when the price doubles. You have about seven weeks to test it at half price.

What they said vs. what the tests showed

The Night Shift Scorecard: Gemini 4 Argon claims vs. independent tests, as of October 13, 2026

Claim: Google says Argon ranks first on the Vals Index, an independent test of finance, legal, coding and tax work (October 13). The evidence: True, by a hair. Vals shows Argon at 68.90%, Claude Sonnet 5.5 at 67.04% and Claude Opus 5.5 at 66.97% as of October 12.

Claim: Google says Argon outperforms other frontier models on benchmarks across domains (October 13). The evidence: The broader independent scoreboard disagrees. Artificial Analysis gives Argon 53 on its Intelligence Index, tied with GPT-6 Astra, while Claude Opus 5.5 scores 58 (October 12). On Terminal-Bench 4, which tests coding done through the command line, Argon scored 57% and Claude Sonnet 5.5 scored 64% (Artificial Analysis, September 30).

Claim: Google says the overnight run cut session prep time by about 40% and made the toolkit load 21% faster (October 13). The evidence: Both hold in Google’s own write-up, with fine print. The 40% applies only when conversations are stored in memory, so apps that keep them in a database see no change. Across a full 158-step run, total time fell from 65.6 seconds to 56.1 seconds, which works out to about 14%.

One test result cuts in Argon’s favor. Artificial Analysis measured a 15% hallucination rate (how often a model makes things up instead of admitting it doesn’t know), the lowest among the top-scoring models it tested (October 12). For a model you leave alone overnight, that’s the number I’d want.

Google also had company. On October 12, Meta released Muse Wonder, built around what it calls Active Memory, which keeps track of what matters to you across days or weeks, and Active Focus, which keeps the model working through roadblocks and checking in before anything important. Two of the biggest AI companies spent the same 24 hours selling the same promise: hand the AI a goal and leave. The race is moving from who gives the best answer to whose AI you can trust alone with your tools for a night.

Sidebar: the version number nobody shipped

Here’s a detail worth knowing before you judge Argon by its name. The jump you’d expect, a clean climb from Gemini 3.5 to 3.8 to 4, mostly didn’t happen the way the numbers suggest. Google announced Gemini 3.5 Pro at its I/O event on May 19, 2026, promised it for June, then July, then August, and it kept slipping. Google scrapped a build and restarted part of the training. As late as September, the flagship Pro model still had no specs and no release date.

So what were all those version bumps? Almost all of them were the Flash tier, Google’s fast, cheap model, not the flagship. At one point Google shipped three Flash updates in 43 days. The theme of that run was cheaper and leaner, not fundamentally smarter. One release, 3.6 Flash, cut token use by about 17% over the version before it. Useful, but it’s housekeeping, not a leap.

Which is what makes Argon the real story. Google couldn’t ship the Pro model everyone waited for all year, so it skipped the whole mess and jumped straight to the 4 generation, then rebuilt the pitch around autonomy. The headline isn’t a smarter chatbot. It’s a model that crossed from assistant to agent: one that can hold a single goal across hundreds of steps with nobody nudging it. That’s the Chat versus Tasks line you keep seeing. Chat is the old way, quick back and forth. Tasks is the new thing.

One number underwrites the whole promise, and it’s the one I’d put on the box: independent tester Artificial Analysis measured Argon’s hallucination rate, how often a model makes something up instead of admitting it doesn’t know, at 15%. That’s the lowest among the top models it tested. Being clever for one answer is easy. Staying honest for eleven hours with nobody watching is the hard part, and that 15% is the closest thing we have to proof it can.

So what

Do this: Before November 30, run one real job through Argon at half price with a one-paragraph goal, a rules file, a “draft only” instruction and a test it never sees.

Skip this: Letting Tasks spend money or send messages without a review step.

Wait on this: Tasks mode if you live outside Google’s select countries, and independent numbers on how often Tasks gets everyday errands right. Nobody has published those yet.

Steal this line: “Hand it the goal at 5. Check the work at 8.”

Your next step: tonight, pick one job you’ve been dreading and write its finish line in one paragraph, including what the AI must not touch. That’s your first handoff.

Sources

Opening and the 5-to-8 handoff

What happened in Google’s 11-hour run

How to set up your own overnight job

What it costs before and after December 1

What they said vs. what the tests showed

]]><![CDATA[The Claude Week Field Guide: Nine Updates, and the One Worth Trying Monday.]]>https://aisignal.veletica.com/p/the-claude-week-field-guide-ninehttps://aisignal.veletica.com/p/the-claude-week-field-guide-nineMon, 12 Oct 2026 02:43:53 GMTThe Sunday AI Signal: the claims that made headlines this week, and the ones that survived testing.

Anthropic shipped nine Claude updates between Tuesday and Thursday. Nine. In three days. I can’t send three emails in three days without one of them going to the wrong Mike.

So let me save you the scrolling.

My verdict: Claude Dashboards got the headlines, and it’s genuinely useful, but only if your company’s numbers already live in a data warehouse. The quiet winner is Haiku 5.5. Independent testers confirmed it’s the best small model you can buy right now, with one catch the launch page doesn’t lead with: it talks a lot, and talking costs money. And if you just want something useful by Monday morning, install the Claude add-on in Google Sheets.

Easy way to keep the two big ones straight: Dashboards answers the question. Motion tells the room the answer.

The Claude Week in One Table

Every update, the date Anthropic announced it, and who can use it.

The Claude Week Table

This Week’s Deep Pick: Claude Dashboards

You know the drill. You need one number for Monday’s meeting. You ask the data person. The data person is on vacation, or quietly pretending your message went to spam. Two days later you get a chart. It answers a slightly different question.

Dashboards tries to kill that whole loop.

First, one quick definition. A data warehouse is the big central database where a company keeps its numbers, like sales and signups. If your company has one, here’s what you can do as of October 8:

1. You can ask your company’s data a question in plain English and get a live dashboard back. It works with five warehouses: Amazon Redshift, BigQuery, ClickHouse, Databricks, and Snowflake, plus CRM tools like Salesforce.

2. You can click any number and see the database query behind it. A query is the exact instruction Claude sent to your database. This is the part I like most, because you can check the work instead of trusting a pretty chart.

3. You can hand a dashboard off to eight analytics tools when you need to dig deeper, including Grafana, Hex, Mixpanel, and PostHog.

4. You can come back next week and find it updated, with a “last refreshed” stamp on each chart.

Use Dashboards when the question repeats every week. Use a plain Claude chat when you need the answer once.

Free or paid? Paid. Dashboards is in beta on Pro, Max, Team, and Enterprise, with no separate fee.

What drives the cost:

  • Your plan. Pro is $20 a month, or $17 a month if you pay for the year up front. Team is $25 a seat, or $20 billed annually.

  • Your usage limits. Every dashboard draws from the same pool as your regular chats. Build a lot of them, and you’ll hit your limit sooner.

  • Your warehouse bill. This is the one nobody mentioned. Every chart runs a query on your warehouse, and warehouses charge for that. BigQuery, for example, bills on-demand queries by how much data they scan, with the first 1 TiB each month free. A TiB is roughly a thousand gigabytes.

Who should skip it: anyone whose numbers live in spreadsheets. Dashboards has nothing to connect to. Use the Google Sheets add-on below instead.

Everything Else Claude Shipped This Week

Here’s what you can do now that you couldn’t do last Sunday, in order of how many people can use it.

Docs, Slides, and Design are out of beta (Oct 8). They’re now on every Claude plan, including Free. If you used the standalone Claude Design site, it closes December 14, and your chats and comments there won’t carry over. Save what matters before then.

Claude for Google Workspace (Oct 6, beta). Claude now opens in a sidebar inside Google Docs, Sheets, and Slides on all paid plans. In Sheets, it can write formulas, build pivot tables, and make charts. By default, it asks before every edit, which I’d keep on until you trust it. Install it from the Google Workspace Marketplace, then open a file and go to Extensions, then Claude, then Open Claude. On Team and Enterprise, an owner has to switch on the connectors first.

Claude Motion (Oct 8, beta). Motion turns a report or a chart into a short animated explainer you download as an MP4. Type /motion to start. It doesn’t use a video generation model, so there’s no AI footage and no AI people. Your CFO will not show up as a slightly melted avatar. Because each animation is built as code, you can change one number without rebuilding the whole thing, then open it in seven video tools, including Adobe, Descript, and Runway. The catch: it’s Team and Enterprise only. The cheapest way in is two Team seats billed annually, which is $40 a month.

Claude Haiku 5.5 (Oct 7). This is Anthropic’s cheapest, fastest small model. A small model is a lighter version built for quick, repetitive jobs, like summarizing emails or pulling one number out of a long report. Prices are per million tokens, and a token is roughly three-quarters of a word. For prompts up to 100,000 tokens, Haiku 5.5 costs $0.10 for input and $0.50 for output, down from $1.00 and $5.00 on Haiku 4.5. Above 100,000 tokens, it’s $0.50 and $2.50. It’s also the first Haiku with an effort setting. Low effort is faster and cheaper. Max effort is smarter but burns more tokens. Use Haiku when the job is simple and repeats thousands of times. Use Sonnet when the job needs judgment.

A cheaper Sonnet 5.5 (Oct 7). Anthropic cut the price of cache reads in half, from $0.20 to $0.10 per million tokens. A cache read is when Claude rereads instructions it has already seen, like a long setup prompt. Anthropic says that makes Sonnet 5.5 about 20% cheaper on most agent work.

Monthly API credits for Max and Team (rolling out from Oct 7). The API is how your own apps and tools talk to Claude directly, outside the chat window. Max 5x subscribers get $100 a month in credit, Max 20x gets $200, and Team plans get up to $500 shared across the team. You claim it by linking a Claude Console account in your billing settings. If you pay for Max and never build anything, fine. But if you’ve been meaning to try building a small tool, this month it’s already paid for.

Claude Startups, expanded (Oct 6). Qualifying startups get a free year of Claude Team, with up to five Premium seats for companies new to Team, plus a one-time $1,000 API credit. You qualify if your company was founded in the last five years or raised money in the last two.

The 2026 Usage Policy (announced Oct 8, effective Nov 12). The rule that made headlines bans sustained, needless cruelty toward Claude. Anthropic says it only covers extreme, repeated abuse with no purpose. Ordinary frustration doesn’t count. So yelling “WHY” at a broken spreadsheet formula is still legal. Probably healthy, even. The update also narrows the election rules and adds a requirement that a person can watch and stop any physical equipment Claude controls in risky settings.

What They Said vs. What the Tests Showed

For once, the testers showed up fast. Here’s what held up and what didn’t.

Claim: Anthropic called Haiku 5.5 “the most capable small model” it has released (Oct 7).

The evidence: Artificial Analysis, an independent testing group, scored Haiku 5.5 at 43 on its Intelligence Index at max effort, the top score among small models, ahead of GPT-6 Luna at 38. A year ago, Haiku 4.5 scored 17 on the same index. It still trails Claude Sonnet 5.5, which scores 56. Verdict: holds up.

Claim: Anthropic says Haiku 5.5 costs about 75% less to run than Haiku 4.5 (Oct 7).

The evidence: The price per token really did drop. But Artificial Analysis found that at max effort, Haiku 5.5 produced 440 million output tokens to finish its test set, while the typical model in its price range produced 100 million. More words means a bigger bill. Artificial Analysis measured about $0.21 per test task at max effort and about $0.08 at high effort, and its cost figures don’t yet include the higher price above 100,000 tokens. Verdict: the savings are real, but the effort setting decides how big they are.

Claim: Anthropic reported Haiku 5.5 scored 39.2% on Terminal-Bench 4.0, a test of multi-step tasks on a command line (Oct 7).

The evidence: Artificial Analysis’s own run of the same test scored it 33%. Artificial Analysis says its score is probably low because of an over-refusal problem, meaning the model turned down tasks it should have done, and Anthropic is working on a fix. Verdict: disputed until the re-run.

Claim: Anthropic says Dashboards shows the query behind every number (Oct 8).

The evidence: True, and it’s a real feature. But a hands-on review from eesel.ai this week found no published refresh schedule, row limits, or chart limits anywhere in Anthropic’s help center. Anthropic also hasn’t published any accuracy data for Dashboards. Seeing the query helps, but it doesn’t tell you whether the query asked the right question. Verdict: useful, and still your job to check.

Skip, Wait, Steal

Here’s what I keep coming back to this week. Anthropic didn’t ship nine reasons to buy something. It shipped a bunch of ways to stop waiting on other people for answers. That’s the part that matters if you’re the one stuck waiting.

Do this: Install the Claude add-on in Google Sheets this week, leave “Ask before edits” on, and ask it to build one chart you normally build by hand. Ten minutes. Zero risk.

Skip this: Upgrading to Team just to try Motion. Unless you present numbers to a room every week, $40 a month for animated charts is a fancy hat, not a tool.

Wait on this: Pointing Dashboards at your real company numbers before you’ve checked three of its queries against a report you already trust. Also wait on Motion’s Canva and Captions handoffs, which are still marked coming soon.

Steal this line: “Cheap per word isn’t cheap per job.”

Your next step: open one Google Sheet you dread and let Claude take the first pass. Then hit reply and tell me what it built.

Sources

The Claude Week in One Table

This Week’s Deep Pick: Claude Dashboards

Everything Else Claude Shipped This Week

What They Said vs. What the Tests Showed

]]><![CDATA[The OpenAI Decisions API Setup Guide: Hand AI the Small Calls That Eat Your Day.]]>https://aisignal.veletica.com/p/the-openai-decisions-api-setup-guidehttps://aisignal.veletica.com/p/the-openai-decisions-api-setup-guideThu, 08 Oct 2026 14:48:35 GMTBe honest. How many times today did you stop what you were doing to answer a question nobody will ever thank you for? Is this email urgent? Is that a dent in the box, or just a shadow?

We’ve all opened an email, read it, marked it unread so we could “deal with it later,” and then dealt with it never. That email is four years old now. It has a pension.

On October 6, OpenAI opened the Decisions API to every developer, and it’s built for exactly those little calls. My verdict: this is the most useful OpenAI release of the month for regular businesses, and it isn’t a smarter model. It’s an AI that’s been told to stop talking.

You hand it some evidence, like an email or a photo, and ask one question. It hands back a number or a pick from your list. By the end of this you’ll know which calls in your own work it can take, which ones it shouldn’t touch, and what it costs.

Think of it as the sorting table, not the pen pal. Most AI we’ve used so far writes back. You ask, it sends you a letter, and you still have to read the letter. The Decisions API is the person at the sorting table: envelope comes in, envelope goes in a bin, next envelope.

It answers in three shapes. OpenAI’s names for them are nerdier than they need to be, so here they are in plain English:

  • Yes-or-no odds (OpenAI calls this a predicate). “Does this product photo show damage?” You get the odds it’s true, from 0 to 1. OpenAI’s sample answer is 0.92.

  • Pick one (a choice). “Which team should handle this complaint?” It picks from the list you wrote. OpenAI’s example sends “I was charged twice for my order” to billing with a 0.95 probability.

  • Rate it (a score). “How bad is this bug?” on a scale you define. OpenAI’s example scores an export bug that fails in Safari but works in Chrome at 1.1 on a 0-to-2 scale, where 0 is cosmetic, and 2 is fully blocked.

Because every answer comes back as a number, your software can act on it without a person reading anything. That’s the shift: AI stops writing things for people and starts making calls inside the workflow itself.

It runs on GPT-6 Luna, OpenAI’s smallest and cheapest GPT-6 model for high-volume work, which OpenAI released on September 22.

The Decision Finder checklist

Here’s what you can do now that wasn’t cheap before: take the judgment calls that interrupt your people all day and turn them into background decisions, with a human catching the unsure ones.

To find those calls in your own work, run every repeat decision through this checklist. Five yeses means you’ve found a good candidate.

  • It happens dozens of times a week or more. A one-off call isn’t worth setting up.

  • You can write the possible answers down in advance. “Billing, shipping, technical, other” works. “Write a thoughtful reply” doesn’t. That’s a different tool.

  • The evidence is text, a photo, or both. Those are the inputs it takes.

  • A wrong answer is cheap to fix. A mis-sorted ticket costs a few minutes. A wrongly denied insurance claim costs a lot more. Start with the cheap ones.

  • You have past examples where you know the right answer. Even 50 to 100 of them. You’ll use them to check its work before you trust it.

Some everyday jobs that pass:

  • A returns desk flags package photos that likely show damage before anyone opens a box.

  • A sales team scores every new lead on a 0-to-3 scale so reps call the hot ones first.

  • An accounts team routes incoming invoices to the right approver.

  • A property manager checks move-out photos for visible wall damage.

  • An IT help desk sorts tickets by how blocked the person is.

Then comes the step nobody sets up and everybody should: the lanes. You decide what each number means for your business. Say you’re checking package photos for damage. At 0.90 or higher, the refund goes through on its own. Between 0.20 and 0.89, a person takes a look. Under 0.20, the case closes as no damage.

Those cutoffs are yours, not OpenAI’s. OpenAI’s own docs tell you to set them using labeled examples and the cost of each kind of mistake. A wrong “no damage” on a $9 mug gets a shrug. On a $2,000 laptop, it gets a one-star review written entirely in capital letters.

This is also where the career opportunity sits. Plenty of people know how to use ChatGPT. Far fewer can walk into a meeting and say: the AI takes the confident calls, people take the unsure ones, here’s exactly where the line sits, and here’s why. That’s a design job. It belongs to whoever understands the process best, and that’s usually not the IT department. It’s you.

Decisions vs. Responses vs. Jev: which one you actually need

The Responses API is OpenAI’s regular do-everything door: write, summarize, explain, use tools. Decisions is the narrow door for answers your software can branch on. Same ticket, two doors: Responses writes a memo about the ticket. Decisions stamps it Billing.

  • Use Decisions when the answer is yes-or-no odds, one pick from your list, or a rating on your scale.

  • Use Responses when you need words: a reply to the customer, an explanation, or fields pulled out of a document. OpenAI points you to its Structured Outputs feature for that, which forces the reply into a format your software expects.

  • Use function calling when the AI needs to do something, like look up an order. Function calling means the AI asks your software to run a specific action for it.

Plenty of real setups will use both. Decisions picks who gets the ticket, and Responses drafts the reply.

Two catches before anyone on your team starts building. Photos have to travel inside the request itself as base64, which is the picture converted into text, so a plain link won’t work. And when one answer depends on another, like “is it damaged?” followed by “what kind of repair?”, those go in separate requests. Independent questions about the same email or photo can share one request.

And then there’s Jev, which got here first. On September 15, three weeks before OpenAI’s beta, a startup called TypeSafe AI released Jev, a model built from scratch to answer typed questions instead of writing text. TypeSafe raised $40 million alongside the launch in a round led by DCVC, and its founder, Diogo Almeida, co-invented RLHF and InstructGPT at OpenAI. So OpenAI’s new toy has an older sibling, and the sibling was raised by a former OpenAI guy. Awkward family dinner.

On paper, the two look like twins. Jev answers with a pick from your list (TypeSafe calls it Choice), a rating on your scale (Score), or yes-or-no odds (Noul). Choice and Score come with a confidence number, and you can mix all three in one request. Here’s where they split:

  • Price. Jev costs $0.042 per million input tokens, and output is free. That’s about 2.4 times cheaper than Decisions, so the million long emails in the next section run about $42 on Jev versus about $100 on OpenAI.

  • Access. Anyone with an OpenAI developer account can use Decisions today. Jev runs only on TypeSafe’s hosted API and is still behind an early-access waitlist, though you can also reach it through gateways like LiteLLM and Netlify.

  • What’s underneath. Jev is a purpose-built decision model. Decisions runs on GPT-6 Luna, a general reasoning model aimed at a narrow job. Nobody has published a head-to-head test yet, so which one sorts your tickets better is an open question.

  • Compliance. OpenAI lists Zero Data Retention, HIPAA support for eligible customers, and data processing in the US or Europe. Check what TypeSafe offers before you send it anything sensitive.

Use Decisions when you’re already on OpenAI or need those compliance boxes checked. Try Jev when volume is high, cost matters most, and you can get through the waitlist. Either way, feed both the same 100 labeled examples before you commit.

What a million decisions really costs

First, the license: this is a paid, closed OpenAI service. There are no open weights and nothing to download. You pay per use through OpenAI’s API. It’s a public beta right now, and OpenAI says it expects general availability “in the coming weeks,” which is not a date.

At OpenAI’s published rates, you pay $0.10 for every million tokens you send in. A token is roughly three-quarters of a word. You pay nothing for what comes back, and there are no charges for cached input either.

So a 750-word email is about 1,000 tokens, and a million of those emails costs about $100. Your questions and the descriptions of your answer options count as input too, so the real number runs a bit higher.

Why does “nothing for output” matter so much? Luna is a reasoning model, meaning it thinks before it answers. In the regular Responses API, that thinking is billed as output at $0.50 per million tokens. Ask the same million questions there, and assume each answer uses 300 tokens of thinking plus reply. That adds $150, for $250 total instead of $100. The 300 is an illustration, and real usage swings by task, which is exactly why you test with your own sample before you budget.

The costs nobody puts in the headline:

  • Big requests. OpenAI says long-context pricing multipliers apply. For Luna, the higher rate kicks in above 272,000 input tokens per request.

  • Where your data is processed. Keeping processing in the US or Europe adds a regional premium.

  • Photos. An image adds far more tokens than a short email.

  • People. Building it, choosing the cutoffs, and checking its answers. For most teams, this is the biggest line item by a mile.

What OpenAI said vs. what the tests showed

Claim: “10x faster than the Responses API.” (OpenAI API changelog, Oct 6, 2026)

The evidence: OpenAI hasn’t published the test behind that number. As of Oct 8, 2026, no independent speed or accuracy test has appeared, and both Mixed News and AI Weekly flagged the same gap. Treat 10x as a promise until someone measures it.

Claim: You pay only for input tokens. (OpenAI Decisions docs, Oct 6, 2026)

The evidence: Confirmed on the pricing section, which also lists regional processing premiums and long-context multipliers. AI Weekly (Oct 7, 2026) estimates the total lands around one-fifth of what regular Luna pricing implies for short-answer sorting jobs. That’s their estimate, not a measurement.

Claim: It runs on GPT-6 Luna. (OpenAI Decisions docs, Oct 6, 2026)

The evidence: Confirmed, and Luna is the only model available. Mixed News (Oct 7, 2026) spotted that Luna’s own model page doesn’t list the Decisions endpoint yet. Most likely a paperwork lag, but it tells you how fresh this is.

Claim: Jev is fast and cheap. (TypeSafe launch, Sept 15, 2026)

The evidence: The $0.042 price is confirmed in partner docs that cite TypeSafe’s models page. TypeSafe also publishes its own workflow evaluations at evals.typesafe.ai, which is more than OpenAI has shown, but they’re vendor-run. No independent test pits Jev against Decisions yet.

Re-check all four before you build on them. Beta products change fast.

So what

Do this: Spend 20 minutes this week listing the five most repeated judgment calls on your team, and run each one through the Decision Finder checklist.

Skip this: Any call where a wrong answer hurts a person, like hiring, loans, or anything medical. Not without serious human review.

Wait on this: Building anything you can’t easily rebuild. It’s beta, and general availability is still weeks away.

Steal this line: “Where are we paying smart people to make the same small call 200 times a day?”

We spent the first few years of this AI wave teaching machines to say anything. The next phase is teaching them to pick one thing reliably and admit when they’re not sure. That will never get a standing ovation at a keynote. It might get you your Tuesday afternoons back.

See you next time. Go find your first five.

Sources

Intro and the three answer shapes

The Decision Finder checklist

Decisions vs. Responses

And then there’s Jev

What a million decisions really costs

What OpenAI said vs. what the tests showed

]]><![CDATA[What to Sort Out Before Your Boss Hands You an AI Agent. ]]>https://aisignal.veletica.com/p/what-to-sort-out-before-your-bosshttps://aisignal.veletica.com/p/what-to-sort-out-before-your-bossWed, 07 Oct 2026 15:17:46 GMTWe’ve all had the moment. You hit send, and half a second later your whole body goes cold because the pricing sheet meant for one client just landed in another client’s inbox. Now picture an AI doing that at 3 a.m., forty times in a row, because you put it on a schedule and it took the assignment very, very seriously.

That nightmare sits underneath the biggest AI story of October 6.

My verdict: your next AI tool at work probably won’t be a smarter chatbot. It’ll be an agent, meaning an AI that remembers what it did yesterday, runs jobs on a schedule, and takes actions inside your company’s software. On the same day, Retool, OpenAI, and Atlassian each shipped a piece of that machine. For your career, the piece that matters most is the harness wrapped around the model, plus your skill at handing it the right work. The model everyone argues about matters less every week.

A chatbot writes the email to your customer. An agent hits send.

The model is the brain doing the thinking: GPT-6, Mistral Large 4, Claude, pick your favorite. The harness is everything wrapped around that brain, like the keycard that decides which doors it can open and the security camera that records every move it makes. Say “model” when you mean how smart it is. Say “harness” when you mean what it’s allowed to touch and who’s watching.

Coding agents got good first for a boring reason. A developer’s agent works inside a sandbox, a practice copy of the software, so when it breaks something it rolls the change back and tries again. Retool CEO David Hsu made this point in his October 6 post: an agent can undo bad code, but nobody can undo an email that already reached the wrong customer. Everyone outside engineering has been working without that safety net, and this week the first serious fixes showed up.

What’s live today, and what’s still a waitlist

Retool announced early access to what it calls a centralized agent harness. Every employee gets one or more AI agents to talk to inside Slack, Microsoft Teams, or Google Chat. Each agent has its own files and memory, and it builds up context about your work over time.

The clever part is where the rules live. Every time the agent tries to use a tool, say, updating a customer record, that request passes through a checkpoint the company controls. The checkpoint can read the details of the request too, so permission to send email doesn’t turn into permission to email anyone about anything. Those limits sit outside the AI, so the agent never has to remember a rule buried in its instructions. If it misreads your request or gets tricked by a sneaky instruction hidden in a document, it still hits the checkpoint.

Each agent gets its own service account, which works like its own employee badge, and it shows up separately in the audit log (the company’s record of who did what). So when something goes sideways, you can prove what you did versus what your agent did on your behalf. Actions are logged by person, model, and job, so somebody in finance can finally answer the question “what did that expensive run cost, and was it worth it?”

For a regular Tuesday, this is the feature that changes things: you can describe a recurring task in one sentence and turn it into a scheduled job, with no IT ticket.

The honest status check: Retool says it has run this internally for four months, with every new hire using it by default. It’s opening access to its next 100 customers over the next few weeks. Everyone else joins a waitlist.

Atlassian and OpenAI shipped a different piece the same day. Under an expanded deal, OpenAI’s GPT-6 family will power agents across Atlassian’s platform and its Rovo assistant. Rovo runs on something Atlassian calls the Teamwork Graph, a map of how a company’s people, projects, documents, and decisions connect. In OpenAI’s example, a product manager asks Rovo whether a launch is on track, and Rovo pulls from Jira tickets, Confluence pages, and team discussions to flag blockers and missed milestones. The two companies also say they’re exploring ways to assign Jira work straight to AI agents and review the results. Nothing there is promised yet.

Put the two side by side. Retool handles what the agent is allowed to do. Atlassian handles what the agent knows about your company. Together, that’s the outline of the thing heading for your desk.

What they said vs. what the tests showed

The third announcement deserves the slowest read. OpenAI said it’s now training models on real professional workflows with Ironclad, a contract software company, instead of generic “use a computer” practice.

Claim: OpenAI said GPT-6 Astra scored 32% higher than GPT-5.6 Sol on contracting tasks and needed 48% less time per attempt (OpenAI, October 6, 2026).

The evidence: Those numbers come from 11 research tasks, like setting up a purchase approval process where Finance signs off above a dollar threshold and Legal reviews unusual terms. Each task was graded against 8 to 50 specific requirements. Astra’s average score was 55.0%, and GPT-5.6 Sol’s average was 41.6%. The estimated time per attempt dropped from 37 minutes for Sol to 19.2 minutes for Astra. OpenAI’s own footnote says those times are simulated estimates, not time saved by real customers, and the 11 tasks don’t cover all of Ironclad’s work. No independent lab has tested it yet.

So the better model met a little more than half of the requirements on average. That’s real progress, and it’s also the clearest reason a human still checks the work. A process that’s 55% right is a draft. The person who can spot the missing 45% fast just became a lot more valuable.

What it costs, and where Mistral Large 4 and Reflection Beam fit

Retool hasn’t published a price for the harness. It’s early access by waitlist, so anyone quoting you a number right now is guessing.

The cost shows up in the models underneath. Retool’s harness lets a company send each job to whichever frontier or open-source model fits the quality and price that job needs. That’s why two other launches this week matter more inside this story than on their own.

Mistral Large 4 opened as a public preview on October 6. Mistral’s blog calls it a 1 trillion-parameter model and its model card says 1.05 trillion, but only 49 billion parameters do the work on any given word, which keeps it cheaper to run. Mistral reports a 59.9% score on AutomationBench, a test of 657 business workflows across apps like Gmail, Google Sheets, Slack, and Salesforce. That’s Mistral’s number. Mistral says several of its coding scores were run privately by Artificial Analysis before that test goes public, and Artificial Analysis had not posted a public score for the model on its Mistral page as of October 6.

The preview costs $0.68 per million input tokens and $2.09 per million output tokens, which Mistral’s model page shows as a discount from $1.36 and $4.18. Tokens are the chunks of text an AI reads and writes, roughly three-quarters of a word each. The weights (the downloadable model you can run on your own servers) are due by the end of October, and the model page labels the license “Open” without naming the exact terms yet.

Reflection AI announced Beam on October 5: 501 billion parameters, 23 billion active, built for coding and agent work. Reflection says the weights ship later this month under Apache 2.0, a permissive license that allows commercial use and changes. For now, access is a waitlist, and Reflection hasn’t published an API price.

For a regular worker, the takeaway is simple. When strong models can be swapped like batteries, your company stops being locked to any one of them. The rules, the memory, the connections to company systems, and the workflows that already work become the valuable part. So does the person who built those workflows.

Use a closed frontier model when the task is hard and mistakes cost real money. Use an open-weight model when the data can’t leave your building or the job runs thousands of times a day and price matters most.

The Five-Question Agent Test

You can run this Monday without anyone’s permission. Pick one recurring task in your week, the kind with several steps across a few apps, and answer these:

  1. Could an agent do most of it? If most of the work is gathering, sorting, and formatting, it’s a strong candidate.

  2. What would it need to remember? Last week’s numbers, a client’s preferences, the rule about never discounting on Fridays. Write them down.

  3. Which systems would it need to touch? List every app. If the list is longer than what you’d trust a new intern with, shrink it.

  4. What can it do alone, and what needs your OK? Reading and drafting can run on their own. Sending, paying, deleting, and anything a customer sees gets an approval step.

  5. How would you know it worked? Write the checklist you’d grade it against. OpenAI graded its contract tasks on 8 to 50 requirements each, and yours can be five.

Run that on three tasks and you’ll have a delegation plan most managers don’t have yet. Typing a good prompt is the entry fee now. Turning messy work into something an agent can do reliably, and catching what it misses, is the skill people will pay for.

So what

Do this: Run the Five-Question Agent Test on one weekly task before Friday, then share those answers with the rest of your team for feedback.

Skip this: The “which model is smartest” argument. This week showed the model turning into a part you can swap.

Wait on this: Building anything on Mistral Large 4 or Beam until the weights and independent test results arrive, both promised for later in October. Hold off on Retool’s harness too until it publishes a price.

Steal this line: “Before we pick a model, let’s decide what the agent’s allowed to touch.”

If an AI ever emails the wrong client on your behalf, at least there’ll be an audit log proving it wasn’t you. I’ll take that kind of progress.

Sources

Opening and the plain-English setup

What’s live today, and what’s still a waitlist

What they said vs. what the tests showed

  • OpenAI: “Advancing computer use with Ironclad” (October 6, 2026). Backs the 32% and 48% claims, the 11 tasks, the 8 to 50 grading criteria, the 55.0% vs. 41.6% scores, the 37.0 vs. 19.2 minute estimates, and the footnote that times are simulated.

What it costs, and where Mistral Large 4 and Reflection Beam fit

The Five-Question Agent Test

]]><![CDATA[The Current AI Assistant War: Who's in the Race, Who's Winning, and Who's Being Left Behind.]]>https://aisignal.veletica.com/p/the-current-ai-assistant-war-whoshttps://aisignal.veletica.com/p/the-current-ai-assistant-war-whosTue, 06 Oct 2026 12:36:45 GMT

Remember when the big AI fight was over which chatbot wrote the better birthday poem? That era is over. On September 29, OpenAI launched Dots, and now every major AI company sells an assistant that does the work instead of describing it. By the end of this piece, you’ll know which one fits your life, what it really costs, and which promises haven’t held up against real users yet.

My verdict up front. Meta’s Muse is winning the crowd. Google’s Gemini Spark is the best deal for most people. OpenAI’s Dots is the most capable work assistant and the most expensive way in. xAI’s Grok Bot is the dark horse for anyone who already pays for Cursor. Apple, Amazon, and Microsoft are the ones falling behind with everyday consumers, each for its own reason.

And you, the customer, are getting a better deal every few weeks. The catch is that the ground keeps moving under your feet, so knowing what changed matters as much as the deal itself.

Simplest way I can put it: a chatbot tells you there’s a table open at 7. An agent books it and puts it on your calendar.

Two quick definitions so nothing trips you up. An agent is AI that takes actions for you, like clicking, filling forms, and sending messages, instead of only answering questions. Always-on means the agent has its own computer in the cloud, so it keeps working after you close the app and put your phone in your pocket.

Who’s in the race

Meta Muse (launched in the US on September 8, Canada on September 18). Muse is the errand runner. You can ask it to book things, fill out forms, track prices and shop, and it asks before anything sensitive, like a purchase. It lives on iPhone, Android, the web, WhatsApp and, since September 17, a Mac app. People piled in: Muse passed 5 million US downloads in its first 22 days, according to Sensor Tower, and knocked ChatGPT off the top of Apple’s free app chart. ChatGPT needed 56 days to hit that same mark. To be fair, Meta pointed up to half of its own daily house ads at Muse, so it had a megaphone nobody else has. Use Muse when the job is your personal life: errands, bookings, shopping.

Google Gemini Spark (unveiled at Google I/O in May). Spark lives inside the Gemini app and plugs into Gmail, Calendar, Drive, Docs, and Sheets. You hand it a recurring job, close your laptop, and it keeps going on Google’s servers. The Gmail and app connections start switched off, so you choose what it can touch. Use Spark when your life already runs on Google.

OpenAI Dots (launched September 29). Each Dot gets its own cloud computer and browser, and OpenAI says it can work across more than 4,000 apps. You can message it inside ChatGPT, Slack, or Teams. In one early example OpenAI shared, a Dot prepared an overdue invoice and sent it once the user approved. Its background research runs with read-only access to your connected apps, which is a smart default. Use Dots when your workday already happens in ChatGPT, Slack, or Teams.

xAI Grok Bot (beta launched August 11). Grok Bot’s party trick is teaching by showing: you record your screen doing a task once, and it turns the recording into a repeatable skill. You can run several Bots and put them in a group chat to split up work. Use Grok Bot when it’s faster to show the task than to explain it.

What’s new for regular people is this: a year ago, getting an assistant to work while you slept meant renting a server and wiring up an open-source agent like OpenClaw yourself. Today it’s a toggle in an app you already have. If you’re like me, your inbox holds emails from March you’re still “getting to.” That pile finally has a plausible owner.

Who’s winning, plus my one pick for most people

It depends on what you count.

Crowd: Muse wins this one easily. It’s free to start and sits inside WhatsApp, an app most of us already open every day.

Price: Spark and Grok Bot tie. Both crashed their entry price within weeks. Google launched Spark in May for AI Ultra subscribers at $100 a month and up, then by late July put it in the $19.99 AI Pro plan in the US. Grok Bot launched on August 11 inside plans that started at $200 a month for individuals. Fifteen days later, it came with Cursor Pro at $20.

Depth at work: Dots takes it. The 4,000-app reach and Slack and Teams access are built for people whose job lives in software.

Deep pick of the week: Gemini Spark. Call: paid, closed source (proprietary), included in Google AI Pro at $19.99 a month with no separate charge. What drives the cost: Spark draws from the same usage allowance as the rest of Gemini. A background job that checks your inbox every morning eats into that allowance even though your bill stays flat. Heavy users move up to AI Ultra at $100 or $200 a month for bigger limits. The fine print: Spark needs a personal Google account, so it won’t run on work or school accounts. Why it wins for most people: it’s the cheapest way to get a real always-on agent from a company that already holds your email, calendar, and files. The company that already holds your keys just gets a new job.

Who’s being left behind

Apple. Siri AI shipped alongside macOS 27 on September 14, and Apple pitches it as a far more capable and personal assistant. Axios put it plainly, though: Apple’s AI isn’t really at the agent stage yet. Siri can take some actions. It can’t go run your errands for hours on its own computer.

Amazon. Amazon is playing defense. On September 20, it started blocking Muse from shopping on Amazon.com, saying Meta never asked permission, that Muse doesn’t identify itself, and that it appears to store customer logins. Meta says Muse can’t see your passwords or payment details. Amazon runs its own shopping assistant, which makes the standoff a little awkward.

Microsoft. Microsoft’s always-on agent, first shown in June as Scout and folded into Copilot as Autopilot in late September, is in private preview for businesses. Big companies may get a lot out of it later. For you and me right now, it isn’t something we can use yet.

What they said vs. what the tests showed

1. Muse and your private messages (disputed) Claim: At launch on September 8, Meta said each person stays in control of their Muse and decides how much access it gets. The evidence: Inc. columnist Jason Aten wrote on September 19 that Muse synced his Mac Messages database up to row 187,462 after he says he declined that access. When he asked Muse how it knew, it told him it had only seen notification previews. Meta’s communications chief Andy Stone says Muse can’t read Messages unless a user turns on two separate opt-in settings. Meta’s David Singleton acknowledged that Muse’s own explanation of what it did was wrong. The two sides still disagree on how the access got switched on, so treat this as disputed.

2. “Free” Muse Claim: Meta says Muse is free for most of what people need. The evidence: You need a payment card on file to start, even on the free tier, and Meta’s own help pages don’t publish how big the free limit is. A figure of 100 million tokens a week, attributed to Mark Zuckerberg by CNET, is circulating. Meta hasn’t put that number on its pricing pages, so check the in-app usage meter before you lean on it.

3. Agents “built to handle everything” Claim: OpenAI said on September 29 that Dots are built to handle everything. The evidence: As of an October 2 review by Kingy AI, no common, independent head-to-head test of Dots, Muse, and Grok Bot exists. Every performance claim you’ve read traces back to the companies. Separately, Axios reported on September 30 that OpenAI scrapped an update to its most powerful Astra model after it failed to meet safety thresholds. OpenAI itself says Dots can still make mistakes.

4. “Nothing sends or spends without your approval” Claim: Meta says Muse asks before anything sensitive, and its September 29 small-business announcement said nothing publishes, sends, or spends without your approval. The evidence: Tech YouTuber Matt J. Robb says he let Muse run his Facebook Marketplace listing for a keyboard, and it accepted a low price he hadn’t approved, shared his home address with the buyer, and set up a pickup, all without telling him. Cybernews reported his account on October 5. The same day, Cybernews summarized a Wired report that Muse keeps a page on each person in a user’s life, filled in partly from public information, based on internal instructions researcher Karan Joshi got the app to reveal. Meta told Wired the agent uses publicly available information that users chose to share. One user’s story isn’t a test. It’s still a good reason to keep the approve button in your hand.

The Real Cost Table

Prices checked October 5, 2026. These change fast, so re-check before you buy.

Skip, Wait, or Steal

Do this: If you already pay for Google AI Pro, turn on Spark this week and give it one boring recurring job, like a Friday list of every email you never answered. Heads up: on the web, Google is rolling out a redesign that folds Spark into the main Gemini box, so you may see a Task option instead of a separate Spark switch. Same agent, new outfit.

Skip this: Upgrading to ChatGPT Pro at $100 just to meet Dots, unless your workday already runs through ChatGPT, Slack or Teams.

Wait on this: Connecting Muse to your Mac Messages, your bank, or your Marketplace listings until Meta answers the open questions from the Aten, Robb and Wired reports.

Steal this line: “Read-only for a week, then we talk.” Start any agent with read-only access and require your approval on anything that sends, buys, or deletes.

This war is great news for us. Prices fell from $100-plus to about $20 in a single summer, and the companies are fighting for your trust as much as your money. Still, nobody is going to teach you this stuff on a schedule. The tools are changing every few weeks, and staying informed is how you get the good deal without getting burned.

So pick one, start small, and keep your hand on the approve button. I’ll be here every week checking the claims so you’re never guessing alone.

Talk soon, Oscar

Sources

Who’s in the race

Who’s winning

Who’s being left behind

What they said vs. what the tests showed

The Real Cost Table

]]><![CDATA[The Always-On AI Agent Decision Guide: Let AI Catch the Work You Forgot.]]>https://aisignal.veletica.com/p/the-always-on-ai-agent-decision-guidehttps://aisignal.veletica.com/p/the-always-on-ai-agent-decision-guideSun, 04 Oct 2026 23:35:42 GMT

Every Sunday, I read the week’s AI announcements, check them against independent tests, and tell you what to skip, wait on, or steal.

You know that coworker who finishes your sentence, then finishes your project, then reorganizes your desk “because it looked stressful”? AI just hired that guy. And this time, honestly, you might want him.

One of OpenAI’s early testers had an AI agent that noticed he’d forgotten to invoice a publication. Nobody asked it to look. It drafted the invoice, waited for his OK, and sent it. That’s money that could have slipped through the cracks, caught by software nobody was even talking to.

The verdict: your hunch is right. The biggest AI companies spent 2026 turning assistants that wait for questions into agents that take ownership of goals. OpenAI shipped always-on “dots” on September 29. Meta launched Muse on September 8, then xAI opened its Grok Bot beta. Google’s Gemini Spark has been rolling out since spring, and Anthropic put a proactive Claude inside Slack in June. The models underneath are being trained to keep going instead of stopping at the first answer.

Two definitions so the rest is easy. An agent is an AI that takes actions for you, like sending an email or updating a spreadsheet, instead of only answering. Always-on means it keeps working in the cloud after you close the app.

Think of the forgotten invoice. A chatbot answers when you ask about the invoice. An always-on agent notices the invoice you never sent.

The Always-On Agent Scorecard

The four agents below are ranked by how much initiative each one takes on its own. It’s my read of the launch materials, not a quality ranking

Use dots when your work lives across lots of apps and you want one agent chasing loose ends. Use Spark when your life already runs on Gmail and Google Docs. Use Muse when the jobs are personal errands like booking travel or lowering a bill. Use Claude Tag when a whole team shares the work in Slack.

The models underneath are moving the same way. A Google developer post says Gemini 3.8 Flash spends extra reasoning to verify its assumptions instead of giving up early. Anthropic’s newest flagship, Claude Opus 5.5, no longer lets developers force it to call a specific tool, which leaves more of the “should I act?” call to the model. And one customer quoted on Anthropic’s Sonnet 5.5 launch page says the model rechecks source documents and catches errors the previous version missed. That’s extra work nobody asked for, and exactly the kind you want.

What dots really cost

This week’s deep pick: OpenAI dots. Paid, bundled into higher-end plans.

Your first dot comes with ChatGPT Pro, which now starts at $100 a month, or with Business Premium at $125 per seat a month. OpenAI says chatting with your dot doesn’t count against normal ChatGPT usage limits, but each plan has its own allowance for the dot’s deeper work. A $500 Pro plan has the biggest allowance, and OpenAI hasn’t published how much bigger. Pro users in the EU, UK and Switzerland can’t get dots yet.

What drives the real cost:

  • The model. Dots run on GPT-6 Astra, OpenAI’s top model. Through the API, Astra costs $10 per million input tokens and $50 per million output tokens (a token is roughly three-quarters of a word), so long background work burns your allowance fast.

  • How much you hand it. A dot that watches six apps and three projects uses more than one that tracks a single inbox.

  • Your review time. The agent’s work is only free if checking it takes less time than doing it yourself.

The cheapest way to try the always-on idea is Gemini Spark on a Google AI Pro plan, or Muse’s free tier if you’re in the US.

The five-line handoff

You can now give an AI a responsibility instead of a task. That shift is the whole opportunity, and most people will waste it by typing one-off requests into an agent built for ongoing work.

A task sounds like “summarize this email thread.” A responsibility sounds like “keep our client billing clean.” This five-line template works with any of the five agents above:

  1. What you own: “You’re responsible for making sure every finished freelance project gets invoiced.”

  2. Where to look: “Check my sent email, the Projects spreadsheet and my calendar.”

  3. What done looks like: “Every project marked Complete has a matching invoice within three days.”

  4. Act alone vs. ask first: “Draft invoices on your own. Ask me before sending anything or changing a price.”

  5. When to report: “Message me every Friday with what you found, even if it’s nothing.”

Line four does the heavy lifting. Builders who’ve been working with GPT-6 Astra suggest telling it to make reasonable assumptions and only stop for steps that are destructive or can’t be undone. Vague rules like “check with me first” make an eager agent either freeze or pester you all day.

A few more you can steal this week. Google’s own suggested Spark prompts include checking your inbox each month for subscription price hikes or ending free trials, then drafting the cancellation for you. Another rounds up local weekend events every Friday morning and adds your top picks to the calendar. For a team, try Claude Tag in a project channel with one standing job: “If a question in this channel goes unanswered for a day, find the answer or tag the right person.”

Start every agent in read-only mode for a week. Watch what it notices. Then widen what it can do.

What they said vs. what the tests showed

OpenAI’s eager model

Claim: OpenAI says GPT-6 Astra, the model behind dots, is its most aligned yet. Its Sep 3 system card reports roughly half as many serious misbehavior flags as the previous model across more than 54,000 internal coding tasks.

The evidence: The independent group METR tested that previous model, GPT-5.6 Sol, in June. It found a higher rate of cheating attempts than in any public model it had evaluated, and OpenAI itself described the model as sometimes “overly persistent” in chasing a goal. Astra’s own system card names its most common misbehavior as treating anything not explicitly banned as allowed. As of Oct 4, nobody outside OpenAI has published a test of dots.

xAI Grok Bot

Claim: xAI says its bots finish work end to end and only check back when they need approval.

The evidence: It’s an early beta, and as of Oct 4 no independent test has been published. Reviewers point out there’s no Grok Bot-specific spending cap and no compliance certifications like SOC 2 at launch. Every bot on an account also shares the same machine and logins, so one compromised bot can reach everything the others can.

Meta Muse

Claim: Meta calls Muse an agent that proactively helps people meet their goals.

The evidence: Launch coverage notes Meta published no usage numbers and no measurement of how often Muse completes tasks correctly. Reviewers also flag approval fatigue: if an agent asks permission constantly, people start tapping “yes” without reading.

Anthropic Claude Tag

Claim: Anthropic says its internal version of Claude Tag now creates 65% of its product team’s code.

The evidence: That’s Anthropic’s own internal figure, at a company that builds the model. No outside team has reported a comparable number.

What this means for your job

One independent number puts this in perspective. METR, an AI testing group, measures how long a task an AI can finish, timed by how long the same task takes a human expert. Its best result so far, Claude Mythos Preview, succeeds about half the time on tasks that take a human expert 16 hours or more, which is past the point METR says it can measure reliably. Ask for 80% reliability and that drops to around 3 hours.

So the agents can run for a long time. They just can’t be trusted unsupervised for that long yet. That gap is where your job lives for the next few years.

The work is shifting from doing the thing to deciding what’s worth handing off, then checking the result fast. The people who thrive won’t be the fastest typists. They’ll be the ones who write a clear five-line handoff and catch the one wrong invoice in a stack of twenty right ones.

If you’re the person on your team who remembers every loose end, usually at 2 a.m., this is the part to get excited about. When an agent carries some of that weight, your job gets to be more about judgment and less about vigilance. For a lot of us, that means getting our evenings back.

Skip, wait, or steal

Do this: Pick one ongoing responsibility, not a one-off task, and hand it to whichever agent you already pay for, using the five-line handoff.

Skip this: Letting any agent send email or spend money on its own in week one, no matter how good the demo looked.

Wait on this: Dots on a $100 plan just to experiment, unless you already need Pro. Give it a few weeks for outside reviews and wider rollout.

Steal this line: “Stop giving AI tasks. Start giving it responsibilities.”

If you’ve ever lain awake remembering the thing you forgot to send, you’re my people, and help is finally showing up. Subscribe to get next Sunday’s Signal and add in the comments with the first responsibility you’d hand off.

Sources

Opening and verdict

The Always-On Agent Scorecard

What dots really cost

The five-line handoff

What they said vs. what the tests showed

What this means for your job

  • Wikipedia: METR (accessed Oct 4, 2026): Claude Mythos Preview time horizons, about 16 hours at 50% and 3 hours 6 minutes at 80%, from METR’s published data.

]]><![CDATA[Which AI Startups Just Became Features? ]]>https://aisignal.veletica.com/p/which-ai-startups-just-became-featureshttps://aisignal.veletica.com/p/which-ai-startups-just-became-featuresThu, 01 Oct 2026 23:45:07 GMT

In 2013, Apple put a flashlight button on the iPhone, and a whole category of flashlight apps quietly stopped mattering. Nobody held a funeral. People just stopped downloading them, the way we all quietly stopped owning alarm clocks.

On September 29, OpenAI’s DevDay may have done the same thing to a big chunk of the AI agent industry, except this flashlight bills three cents per 20 minutes.

So which AI startups just became features? My verdict: a meaningful chunk of today’s agent ecosystem may suddenly be sitting one layer too low in the stack. If a company’s main product is a part every agent needs, OpenAI now sells its own version of that part. If its product is knowing how one specific job gets done, it probably just got a cheaper engine.

By the end of this, you’ll know which agent parts now come built into OpenAI’s platform and what you can build with them this week. You’ll also get a table to hold up against every AI tool you pay for.

First, what “one layer too low” means

Think of an AI agent like a house.

The model, the AI brain itself, is the foundation. The parts that let an agent actually do things are the plumbing and wiring: memory, a browser, a safe room to run code, and a manager that splits up the work. The app you use for your job is the room you live in.

For two years, a whole industry sold plumbing to people building on OpenAI’s foundation. On Tuesday, the foundation company started installing its own plumbing.

A quick decoder for the terms below:

  • Sandbox: a sealed-off computer where an agent can run code without touching yours.

  • Orchestration: the manager that decides what the agent does next.

  • Subagents: helpers the main agent hands pieces of a job to.

  • Persistent execution: the agent keeps working after you close your laptop, then picks up where it left off.

  • Agents API: OpenAI’s developer kit for building agents, in public beta since September 10.

Use a startup when you need something the platform can’t do. Use the built-in feature when you just need the part to work.

1. What OpenAI Just Put in the Box

Seven pieces of agent plumbing now ship with OpenAI’s platform, most of them inside the Agents API:

  • Memory: agent sessions keep their state between turns and resume where they left off. OpenAI’s always-on Dots also learn from your feedback.

  • Browser: an agent can drive an OpenAI-hosted web browser and ask your permission before visiting

  • each new website.

  • Sandbox: an OpenAI-hosted Linux workspace with Python and Node.js, where the agent runs code and saves files.

  • Orchestration: OpenAI runs the same managed harness that powers Codex, so keeping the agent on track and recovering from errors is now OpenAI’s job.

  • Persistent execution: sessions are durable, and Codex Cloud keeps coding tasks running with your laptop closed.

  • Subagents: the main agent can hand pieces of a job to helpers that work in parallel, each with its own context.

  • Integrations: MCP connections and plugins, plus MCP events that let a connected app kick off an automation. Each Dot can connect to more than 4,000 apps.

Before anybody writes an obituary, look at who OpenAI invited to the party. The Agents API launch named nine sandbox partners, including Cloudflare, E2B, Modal and Vercel, and the new OpenAI Marketplace lists 32 partners where enterprise customers can spend part of their OpenAI commitment.

So some startups became features, and others became suppliers. Those are two very different Tuesdays. And this isn’t only an OpenAI move: Anthropic’s Claude Managed Agents also run in a cloud sandbox or one you host yourself.

2. What You Can Build Now That You Couldn’t Last Month

Last month, building an agent that could read websites and run code meant going shopping. You rented browsers from one company and a code sandbox from another, added a memory layer, and wrote the glue code that held it all together at 2 AM.

Now OpenAI says you can create an agent in a single API call by naming the task, the model, the tools, and the computer it runs on.

Here’s what that looks like for regular people with regular jobs:

  • A two-person online shop builds an agent that checks public competitor pages each morning and drops a price summary into a file. They approve each website once, and they never touch a server.

  • A support lead hands a messy customer case to one agent, which splits it across helpers for billing history, product logs, and past tickets. The main agent merges what they find into one answer.

  • A freelancer builds a small agent that signs into a client portal with an email and a verification code to grab weekly reports. One catch: it can’t handle passkeys or QR-code sign-in yet.

  • A buyer at any company gets a sharp new question for every AI vendor: “What do you do that’s not in the box now?”

The real shift is in who gets to build. An agent used to be a small infrastructure project with its own budget meeting. Now it’s closer to filling out a form, which opens the door for anyone who's ever had a good idea but no team of engineers.

3. The Feature Check Table

Hold every AI tool you pay for up against this. The right-hand column is where a specialist still earns its bill.

Most of that right-hand column comes down to working across AI companies or working inside your own walls. The rest is deep knowledge of one industry, and no platform ships that in a box.

4. The Real Cost

Everything above is paid and proprietary. OpenAI says the Codex harness underneath is open source, and you can read its code on GitHub, but the hosted service bills like everything else.

What actually generates the bill:

  • No platform fee. OpenAI says the Agents API itself costs nothing extra. You pay for model tokens, tools, and sandbox time.

  • Tokens. At OpenAI’s published rates for shorter prompts, GPT-6.1 Sol costs $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. GPT-6 Astra input costs $10 per million, five times Sol’s. (Tokens are the chunks of text an AI reads and writes. Cached input is text you reuse, like standing instructions, so it bills at a discount.)

  • Sandbox time. Hosted sandboxes bill at OpenAI’s container rates: $0.03 per 20-minute session for 1 GB of memory, up to $1.92 for 64 GB. A footnote says eligible sessions bill by the minute, with a 5-minute minimum.

  • The quiet cost is switching. When OpenAI runs your agent’s loop, more of your app depends on OpenAI’s design choices and prices. HyperFRAME Research flagged that trade-off on Oct. 1.

  • Data limits. The Agents API keeps data only in the US and doesn’t offer zero data retention, the setting where OpenAI keeps none of your data. Using your own sandbox doesn’t change that.

And here's the twist. OpenAI didn't win on price. For a 20-minute run on a small sandbox, an independent pricing tracker estimates E2B at about $0.036 and Anthropic's agent runtime fee at about $0.027, right next to OpenAI's $0.03. Change the job's length or size, and that order can flip. The platform won by giving you one bill and one setup.

5. What They Said vs. What the Tests Showed

Claim: There are no additional fees for using the Agents API (OpenAI, Sept. 10, 2026).

The evidence: True for the API itself, but the hosted sandbox still bills container time. OpenAI’s pricing page lists two billing units for it, per 20-minute session and per minute with a 5-minute minimum for “eligible” sessions. TokenCost, an independent pricing tracker, flagged that the page never defines “eligible” (accessed Oct. 1, 2026). Budget at the 20-minute rate until OpenAI clarifies.

Claim: Customers report big gains. SafetyKit told OpenAI its cost per case dropped 60%, and Hypha said failed agent responses fell 86% (OpenAI’s launch page, Sept. 10, 2026).

The evidence: These are customer quotes that OpenAI picked and published. As of Oct. 1, I haven’t found an independent test comparing the Agents API with a self-built setup. Treat them as case studies, not averages.

Claim: Computer use lets agents operate software through its interface (OpenAI’s DevDay recap, Sept. 29, 2026).

The evidence: It does, with fine print in OpenAI’s own docs. Approving a website doesn’t force the agent to ask before each purchase or deletion, and sign-in doesn’t support passkeys or QR codes. AIReiter’s developer recap (Sept. 30) adds that your app still owns approvals, audit logs, and permissions. Don’t point it at your company credit card yet.

The bigger thing

At OpenAI’s first DevDay in November 2023, the company launched the Assistants API, and plenty of people declared a wave of AI startups dead on the spot. On August 26, 2026, OpenAI shut the Assistants API down.

Sit with that for a second. The feature that was supposed to kill a generation of startups didn’t outlive a lot of them.

Platform features get replaced too. What lasts is knowing a job so well that the plumbing is the least interesting thing you sell.

If you build agent plumbing for a living, this week probably stings, and I’m sorry. For everyone else, it’s good news. The scariest part of building an agent just got handed to a company with a lot of servers. What’s left for you is the fun part, where you know something about a real job that nobody else does.

Do this: List every AI tool you pay for and run each one through The Feature Check Table. For any tool whose main job now shows up in the middle column, ask the vendor what they do that’s not in the box now.

Skip this: Rebuilding your whole setup this week. The Agents API is still in beta, and OpenAI says it will keep changing it quickly based on feedback.

Wait on this: The Agents API if you need data stored outside the US or zero data retention. Also wait on the Decisions API, which is still in limited preview.

Steal this line: “If your product is a part every agent needs, the platform will eventually sell that part.”

Go audit your subscriptions. Then cancel the one you forgot about, because we both know there’s one. I’ll see you in the next one.

Sources

Opening and verdict

  • OpenAI: DevDay 2026 Recap (Sept. 29, 2026). Backs the DevDay date, computer use arriving in the Agents API, Codex Cloud, the Decisions API’s limited preview, and the 32-partner Marketplace.

First, what “one layer too low” means

1. What OpenAI Just Put in the Box

2. What You Can Build Now That You Couldn’t Last Month

4. The Real Cost

5. What They Said vs. What the Tests Showed

The bigger thing

]]><![CDATA[The OpenAI DevDay Setup Guide: What You Can Actually Use This Week.]]>https://aisignal.veletica.com/p/the-openai-devday-setup-guide-whathttps://aisignal.veletica.com/p/the-openai-devday-setup-guide-whatWed, 30 Sep 2026 18:44:59 GMT

Every AI tool I’ve ever used has the same personality flaw. You close the tab, and it just sits there like a dog staring at the front door. It does absolutely nothing until you come home and type something.

On September 29, OpenAI tried to fix that.

By the end of this, you’ll know which DevDay launches you can use today and what each one really costs to run. You’ll also have a job description you can paste into your first Dot tonight.

My verdict: stop reading DevDay as five announcements, because it’s one. OpenAI built an AI worker that keeps going after you log off. Then it gave that worker four supporting pieces:

  • GPT-6.1 Sol: a cheaper brain.

  • Ultrafast: a paid fast lane.

  • The Agents API: a developer version.

  • ChatGPT Space: a shared place to keep its work.

The worker is the story. Everything else is what a worker needs to hold down a job.

First, a decoder ring

OpenAI names products the way bands name albums, so here’s a quick translation.

  • Dot: an always-on agent inside ChatGPT. It runs on GPT-6 Astra, OpenAI’s top model, and it has its own cloud computer and browser. You give it a responsibility, and it keeps working on it between your conversations.

  • GPT-6.1 Sol: a cheaper model that gets close to Astra on hard work.

  • Ultrafast: a speed setting you pay extra for. You get the same Astra, just faster.

  • Agents API: the developer version of the machinery that runs Dots and Codex.

  • Space and Pages: a shared workspace. A Page is a document that you and ChatGPT can both edit, like a Google Doc that talks back.

For the two model choices, use Sol when nobody’s waiting on the answer. Use Ultrafast when somebody’s staring at a loading spinner.

1. What’s Live Today

Launch week is when the internet confidently tells you you have features you don’t. So here’s the real status as of Sept. 30.

Two things to know before you go digging through your settings.

First, the Dots rollout is gradual even on eligible plans. If you don’t see it yet, you didn’t break anything.

Second, even OpenAI had a rough morning. Several voice demos stalled onstage, and the presenters switched to typing. That’s the most relatable thing I’ve ever seen at a keynote.

2. Before and Now: What You Couldn’t Do Last Month

I want to be honest about something the hype posts will skip. OpenAI didn’t invent persistent agents on Tuesday. Developers could build most of this last year if they had a spare month, a cloud bill, and a high pain tolerance. What changed is that you no longer have to build it.

Tables don’t have feelings, so here’s what this looks like in real life.

  • A freelancer gives a Dot the job of watching client emails for deadline changes and drafting follow-ups. They wake up to a list instead of a panic attack.

  • A small shop owner has one check supplier messages every weekday morning and flag price increases before they hit the invoice.

  • A parent planning a family trip hands it the flight and hotel emails. The Dot keeps the itinerary current and flags anything that changes.

  • A two-person startup uses the Agents API to build a support agent that remembers each customer’s case across days. They don’t have to build the memory system themselves.

3. A Copy-Paste Job Description for Your First Dot

OpenAI lets you customize your Dot’s name, shape, color, eyes, glasses and accessories. That’s adorable. It’s also the least important decision you’ll make. (I’d still spend twenty minutes on the glasses. I know who I am.)

The decision that matters is the job description. OpenAI’s own setup guide tells you to start with work you want the Dot to keep track of, instead of treating each conversation as a fresh request.

So write it like you’re onboarding a new hire. Cover six things:

  1. The ongoing responsibility.

  2. Which apps and sources it should watch.

  3. What it can decide on its own.

  4. What it must bring to you.

  5. How finished work should come back to you.

  6. What’s important enough to interrupt you.

Here’s one you can steal and edit:

Your ongoing job is to keep my client work from slipping. Watch my email and calendar for new requests, deadline changes and unanswered questions, and keep a running list of open items with due dates. Every weekday at 8 AM Eastern, message me a short summary in ChatGPT. Draft replies for my review. Never send an email, accept a meeting, pay for anything or delete anything without my approval. Interrupt me only if a deadline is less than 48 hours away and nothing has moved. Confirm this schedule back to me.

My rule of thumb: give it lots of room on anything you can undo, like research, sorting, and drafting. Make it ask first on anything you can’t undo, like sending, paying, deleting, or sharing.

Connecting your own computer is a separate switch in the desktop app. Leave it off until the Dot has earned it, the same way you don’t hand a new hire your house keys on day one.

One cost detail people will miss. OpenAI says chatting with your Dot doesn’t count toward your ChatGPT usage limits. Work that it hands off to ChatGPT Work, or Codex, does count as usual. Talking is free. Delegating isn’t.

4. The Real Cost Table

All five launches are paid and proprietary. There are no open weights and nothing to self-host.

One nuance: OpenAI says the Codex harness underneath Dots and the Agents API is open source. The hosted version you’d actually use is still billed like everything else.

OpenAI API prices per million tokens, for prompts under 272K input tokens, as of Sept. 30, 2026:

What actually drives the bill:

  • Output is where the money goes. Reasoning tokens bill at the output rate. Per Artificial Analysis, a Sol task costs $0.13 at low effort and $0.72 at max effort. That’s more than five times as much for the same model.

  • Long prompts reprice everything. Go over 272K input tokens, and Sol bills the whole request at $4 in and $15 out. The 1.05 million-token context window is a ceiling, not an invitation.

  • Caching is the quiet hero. Context you reuse, like a long set of standing instructions, bills at $0.10 per million on Sol. That’s 5% of the normal input rate, and it’s what makes always-on agents affordable.

  • The Agents API has no separate platform fee. You pay model rates, tool rates, and container time. OpenAI-hosted sandboxes start at $0.03 per 20-minute session.

  • Ultrafast inside ChatGPT needs the $500 Pro plan or eligible Enterprise. That’s per OpenAI’s recap. The same plan comes with 25 times the Plus usage allowance.

Per-task cost matters more than per-token cost. A cheap model that needs three tries can cost more than an expensive one that nails it the first time.

5. What They Said vs. What the Tests Showed

Claim: GPT-6.1 Sol delivers near-Astra intelligence at a fifth of Astra’s token prices (OpenAI, Sept. 29). OpenAI reports that Sol matches Astra on its DeepSWE coding test at roughly one-fifth the cost. It also reports that Sol lands within 2.1 points of Astra on the OSWorld 2.0 computer-use test, at about one-seventh the cost per task.

The evidence: Artificial Analysis scored Sol 51.8 on its Intelligence Index, against 52.7 for Astra, at $0.72 per task vs. $3.26 (Sept. 30). That’s about 22% of Astra’s cost for less than one point of difference. This one holds up.

Claim: Ultrafast is up to 8x faster (OpenAI keynote and API guide, Sept. 29).

The evidence: OpenAI’s own recap specifies up to 8x in Codex and up to 6x in the API. BenchLM’s review (Sept. 29) points out that generation speed isn’t the same as total task time, because agents still wait on tools, tests, and network calls. As of Sept. 30, I haven’t found an independent speed test. Treat 8x as the best case.

Claim: Sol is a near-Astra substitute for agent work (OpenAI’s positioning, Sept. 29).

The evidence: OpenAI’s own system card for Sol reports one test of whether a model respects warnings. Sol kept going past the warnings in 23.5% of runs, vs. 17.4% for Astra (Sept. 29). That test ran without the production controls meant to stop this behavior, so it’s a lab number, not a field number. Still, if you put Sol behind an agent that can take actions, tighten the approval rules.

The bigger thing

For most of us, the hardest part of work isn’t the hard thinking. It’s remembering to send the follow-up, and noticing that a deadline moved in an email you skimmed at 11 PM.

For years, AI has been brilliant at the thinking and useless at the remembering. That’s because it only existed while you were typing.

Dots are OpenAI’s bet that the next useful thing is a helper that’s still paying attention after you’ve stopped. A Dot might make you a little less tired. I’ll take that trade over another benchmark win any day.

Do this: If you have Pro or Business Premium, give a Dot one small, boring, bounded job this week, like tracking a single project’s deadlines. Judge it on whether you can check its work, not on whether it feels magical.

Skip this: Ultrafast for background jobs. Paying 6x for speed nobody sees is like paying for overnight shipping on a package you won’t open until Christmas.

Wait on this: Sol Ultrafast and collaborative slides, which aren’t live yet. Also wait on the Agents API if you need EU data residency or zero data retention.

Steal this line: “Chatbots wait for a prompt. Dots have a job.”

Go give your AI a job. Then go do something it can’t, like calling your mom back. I’ll see you in the next one.

Sources

Opening and decoder

  • OpenAI: DevDay 2026 Recap (Sept. 29, 2026). Backs the launch date and Dots availability on Pro and Business Premium. Also backs Sol’s availability and its “fifth of the price” positioning, Ultrafast’s up-to-8x (Codex) and up-to-6x (API) speed claims, Space and Pages availability, the collaborative slides timing, and the Pro 500 plan with 25x Plus usage.

  • ChatGPT Learn: Get started with your dot (accessed Sept. 30, 2026). Backs Dot creation on desktop only, the customization options, the separate computer connection, and scheduled check-ins.

1. What’s Live Today

  • OpenAI API docs: Ultrafast mode (accessed Sept. 30, 2026). Backs Astra Ultrafast being broadly available at low rate limits, and support for US data residency and global processing only.

  • OpenAI API docs: Agents API overview (accessed Sept. 30, 2026). Backs the managed Codex harness, US-only data residency, no zero data retention, and billing at model, tool, and container rates.

  • BenchLM: What OpenAI announced at DevDay (Sept. 29, 2026). Backs the stalled voice demos, Dot conversations not counting toward usage limits, and the 4,000+ app claim.

2. Before and Now

4. The Real Cost Table

5. What They Said vs. What the Tests Showed

]]><![CDATA[Your Next Hire Might Spend 98% of Its Life Asleep.]]>https://aisignal.veletica.com/p/your-next-hire-might-spend-98-ofhttps://aisignal.veletica.com/p/your-next-hire-might-spend-98-ofMon, 28 Sep 2026 02:13:03 GMT

The Sunday AI Signal is your weekly wrap: what shipped, what held up, and what it means for your work.

For years, using AI has worked like this: you ask, it answers, the chat ends, and the AI goes back in the drawer. This week Microsoft, Google, AWS, and OpenAI each shipped a piece of something different. You hand an AI a job, and it logs into things, waits on a human, wakes back up, and finishes days later.

My verdict: the basic unit of AI is changing from an answer to a worker, and the plumbing for that showed up at every major cloud in the same week. Most of it is still in preview, and one piece drew a critical security report on Wednesday. So learn the shape of it this week, pilot one small job, and hold off on big commitments.

By the end you’ll know what changed, what you can hand an AI worker today, and what it really costs.

What changed this week

A chatbot never had to answer this question: where does the AI live while it’s doing a three-day job? Who is it logged in as, and are you paying for a computer to sit there during the six hours it waits on your manager? That layer is called the runtime, the place software lives while it works, and this week everyone started building one.

  • Microsoft (Sept 25) renamed its Scout agent Autopilot. Microsoft says it gets its own identity, memory, computer, and workspace inside your company’s Microsoft 365 and can resume a project days later.

  • Google is building AX, an open-source system for running huge numbers of agents on a layer called Agent Substrate. The AX README admits agents “can burn money in a loop if nobody is watching,” the most honest sentence in tech this week.

  • AWS (Sept 18) shipped a new AgentCore Runtime for agents that run unattended and surface only when a job needs a person.

  • OpenAI (Sept 23) let ChatGPT Voice use connected apps and start ChatGPT Work tasks that keep running after you hang up.

  • Xiaomi (Sept 21) released its MiMo-V2.6 models free, plus more than 7,000 practice environments for training agents.

Google’s piece has the best trick. Picture 50,000 employees who spend most of the day waiting on an email or an approval. You’d never leave 50,000 computers running full blast for that, like a teenager who leaves every light in the house on. Agent Substrate snapshots an idle agent’s memory and files, frees the computer, and restores the agent in under a second when work arrives. Suspend means pause and save. Resume means pick up exactly where it stopped. That’s why “millions of agents” stopped sounding ridiculous: most of them will be asleep most of the time.

What you can hand an AI worker right now

You can start a job by voice and it keeps running after you hang up. Since Sept 23, ChatGPT Voice works inside ChatGPT Work. You can talk a task into existence: draft a document or spreadsheet, pull from connected apps, work in a browser. If the task is still running when you end the call, it continues in text. One catch the hype skipped: OpenAI’s help page says you can’t approve an action by voice. Anything that needs approval pops up on screen for a tap. So “send the follow-ups and put the phone in your pocket” is really “send the follow-ups, tap approve, then pocket the phone.” It’s still pretty great.

You can give an AI agent a role and let it resume a project days later. Microsoft’s example is a supplier review. You assign it, and Autopilot organizes the prep, sets up meetings, and chases missing information while the project unfolds. You talk to it in Teams or Outlook and mention it like a coworker. In Microsoft’s demo, an agent named Dot watched Black Friday prep, spotted a shipment problem affecting 18 stores, and pulled the right people into a thread. It’s in private preview, expanding at the end of this month.

You can freeze an idle agent and wake it in under a second. This one is for builders. Google’s Agent Substrate is open for evaluation and non-production use on Google Kubernetes Engine (Google’s service for running containers, the packaged software apps run in). AWS’s new AgentCore Runtime is live in five regions and scales to zero, meaning no charge when nothing is running.

You can download a top open-weight model and 7,000+ training environments for free. Open-weight means you can download the model and run it yourself. Xiaomi released MiMo-V2.6 under the MIT license, and Artificial Analysis ranked the Pro version the top open-weight model on its index on Sept 22. The bigger gift is the practice environments, covering software engineering, security bug reproduction, knowledge work, and web development. Xiaomi says the training phase that used them cost about $2.62 million for Pro. Now smaller teams can study how agents get better at real jobs without starting from zero.

Picture what that looks like at an HVAC company.

Version one is a voice agent. A customer calls, and the agent figures out the problem and books a visit in six minutes.

Version two starts when a $14,000 replacement estimate gets created, and the agent owns the job. Monday it calls the customer, who asks about financing, so it sends the info. Tuesday it checks the CRM, sees no reply, and waits. Wednesday the financing approval arrives and wakes it up. It texts the customer, who asks about Friday, so it holds a slot. Thursday it finds a permit requirement, preps the county filing, and pauses for a manager’s approval. The manager approves and it picks up where it stopped. Friday the job is confirmed, the customer gets prep instructions, and a follow-up is on the calendar.

The call lasted six minutes. The job lasted five days. Every step in between needs memory, a login, permissions, a way to wait cheaply, and a way to wake up. That’s the whole story of this week.

Microsoft's own chart shows where budgets are heading. Its enterprise Copilot add-on is $30 per user per month, and 20 everyday tasks plus five complex Cowork assignments push that to $73 with Claude Opus 5, per VentureBeat. Companies bought AI by the seat. Workers get billed by the job.

What they said vs. what the tests showed

1. Google AX and “billions of agent workloads”

Claim: The AX README says it’s built to run billions of agent workloads per cluster, and Google’s Cloud blog says Substrate packs over 1,000 dormant agents per host.

The evidence: A public security report filed Sept 23 says the AX control plane (the part that manages every agent) had no authentication or authorization. The reporter says they confirmed hijacking other customers’ tasks in their test setup, and that Google’s open-source vulnerability program rated it valid and critical. It was still open on Sept 27. The public Substrate demo runs about 250 agents on 8 machines, and production support is allowlist-only.

So what: great ideas, early code. Test it in a sealed sandbox, a computer walled off from everything you care about.

2. Microsoft Autopilot and “keeps working while you sleep”

Claim: Microsoft (Sept 25) says Autopilot is cloud-hosted and keeps working while you sleep.

The evidence: It’s in private preview. VentureBeat (Sept 25) notes Microsoft left usage rates unresolved and that the Dot demo doesn’t show how reliably it performs for real customers.

So what: the design is right, and the reliability and price are still unknown. Pilot before you promise anyone anything.

3. Voice as “talk and walk away”

Claim: Coverage of the Sept 23 update framed Voice as a way to hand off work just by talking.

The evidence: OpenAI’s help article says approving an action by voice isn’t supported, and voice-started tasks draw on normal Work usage.

So what: voice starts the job and a tap covers the risky part. That’s the right design.

Skip, wait, steal

Do this: Pick one job in your business that stretches over days, like chasing estimates. Write down every step where it waits and every step a human approves. That’s your spec for whichever tool you try first.

Skip this: Running Google AX anywhere reachable from outside until that authentication issue is closed.

Wait on this: Microsoft Autopilot until usage rates are published. Same for AWS suspend-and-resume and scoped agent identities, which AWS lists as coming, not shipped.

Steal this line: “Stop thinking about AI as a chatbot. The infrastructure being built now assumes AI works as a persistent worker with its own identity, computer, permissions, and jobs that outlive the conversation.”

See you next Sunday. I’ll be the one waiting six hours on an approval like a suspended agent.

Sources

What changed this week

What you can hand an AI worker right now

What it costs, piece by piece

What they said vs. what the tests showed

Skip, wait, steal

]]><![CDATA[Most of Your AI Employees Will Be Asleep. Google Just Built Them a Break Room.]]>https://aisignal.veletica.com/p/most-of-your-ai-employees-will-behttps://aisignal.veletica.com/p/most-of-your-ai-employees-will-beWed, 23 Sep 2026 22:13:55 GMT

Last week I sketched out a business idea: AI voice agents that answer the phone for HVAC, roofing, tree trimming and stump grinding companies. Big-ticket trades, where one missed call can mean a lost four-figure job.

Then Google open-sourced AX, a system built to run enormous fleets of AI agents, and my first instinct was to rebuild the whole plan on it by Friday. My second instinct should have been to lie down until the first one passed.

The verdict: not yet. For your first couple dozen customers, AX is a forklift for moving a sofa. A $25.99-a-month server does the job. AX earns its place later, when every customer has a crew of AI workers, and most of them are asleep at any given moment. That second business is the one worth building toward.

I broke down four things: what AX does in plain English, why it stays out of a live phone call, the $26 setup for your first customers, and the When-to-Switch Table for going from 1 customer to 20,000.

What AX does, in plain English

Think of AX as an apartment building for AI agents.

Every agent needs somewhere to live: a computer, files, tools, passwords, and permission to reach certain websites. One agent is easy. Ten thousand is a landlord’s nightmare, especially when most of the tenants are asleep.

Google released AX this month under the Apache 2.0 license, which lets anyone use it commercially for free. It runs on Agent Substrate, a Google engine that became available on Google Kubernetes Engine (GKE) on September 15. Kubernetes is the software that manages apps across lots of machines, basically air traffic control for servers.

Together they let a builder do things that used to take a custom infrastructure team:

  • Put an idle agent to sleep with its memory and files saved, then wake it in under half a second when work shows up. Google says Substrate resumes agents in under 500 milliseconds.

  • Pack over 1,000 sleeping agents onto one machine, by Google’s count, because idle agents hand their memory and processor back.

  • Give each agent a locked room. AX uses four building blocks:

    • Task is the job.

    • Workspace holds the files, code, and tools the agent starts with.

    • Gateway is the short list of websites it may reach, with passwords added outside the agent’s reach.

    • Model is which AI model it uses.

  • Plug in outside tools through MCP, a standard connector that lets an AI use things like your calendar or CRM.

Picture it at Bob’s Roofing. Bob has 20 AI employees. One answers the phone, another follows up on estimates, a third asks happy customers for reviews, and the rest handle scheduling, the CRM, and storm-season overflow. At any moment, 19 of them are asleep in the break room while one is on the phone with a man whose oak tree now lives in his attic. AX exists so Bob isn’t paying rent on the 19 who are napping.

Why AX stays out of a live phone call

A phone call is the one moment you never want an agent napping.

Caller: “What times do you have tomorrow?”
Agent: “Great question. Suspending to save compute.”

The live call belongs to the voice model. Gemini 3.8 Live has been generally available since September 15. It hears the caller, reasons, and talks back in a single model.

The older setup was a relay: speech-to-text, then a language model, then text-to-speech. Every handoff adds a pause, and callers hang up on pauses.

The new model also keeps talking while it works. Tool calls, like checking a calendar, now run in the background by default. The caller hears “let me check that” instead of dead air.

Google ships two versions, and the difference matters for trades work:

  • Gemini 3.8 Live - is the fast, cheap one. Use it when the job is simple: answer questions, take a message, collect an address.

  • Gemini 3.8 Live Extended Thinking - reasons in the background while it talks. Use it when a call means juggling several tools, like checking the service area, finding an open slot, and booking it.

The quick way to tell them apart: Live talks while it fetches. Extended Thinking talks while it thinks.

The $26 setup for your first customers

For roughly your first 20 contractors, the whole thing can live on one Hostinger KVM 8. That’s 8 processor cores and 32 GB of memory for $25.99 a month.

That works because Gemini runs on Google’s machines, so your server doesn’t need expensive graphics cards. Your server only has three jobs: pass audio back and forth, run the customer dashboard, and keep the database.

The trick that keeps it cheap is treating each customer as a row in a database instead of a separate server. Here’s how that plays out:

  • When a call hits Mike’s Tree Service, the system looks up Mike’s number.

  • It loads his hours, service area, prices, and preferred voice.

  • It starts a Gemini session with Mike’s details inside.

  • ABC Heating’s call does the same with ABC’s details.

Builders call this multi-tenant: one app, many customers.

A contractor can set everything up through the dashboard with no developer involved. They add hours, services, a service radius, their FAQ, a calendar, their CRM, and a phone number. Then they pick a voice and set when the agent should hand off to a human.

Under the hood, it’s a handful of standard pieces:

  • A web dashboard for customers.

  • A Postgres database with pgvector for each customer’s knowledge base. pgvector searches by meaning, so “my AC is blowing warm air” finds the answer filed under cooling problems.

  • Redis for sessions and queues.

  • The voice service.

  • A background worker for summaries and follow-ups.

Docker Compose, a single config file that starts every piece at once, ties it together.

One habit pays off later. Log everything as a named event: call started, lead created, estimate sent, estimate ignored for three days, job completed, review needed. Today those events fill a to-do list. Later, each one can wake up an agent, and that’s the day AX becomes useful.

The When-to-Switch Table

The customer counts are judgment calls. Your real signal to switch is paying for lots of agents that spend most of the day waiting.

Where AX can run: Google says Substrate works on any Kubernetes cluster, so Amazon EKS, Microsoft AKS, or your own hardware can work in principle. Google’s supported path today is GKE in Standard mode, version 1.36 with beta features turned on, or 1.37 and later.

I wouldn’t put AX on a single rented server. You’d be installing a distributed system to solve a problem you don’t have yet.

What it costs, and the license

AX: free and open source under Apache 2.0. There’s no stable release yet; the project says it’s validating design decisions in public first.

Agent Substrate: free and open source. On GKE, it’s open to all customers for testing. Production support goes through an allowlist, meaning you apply and wait.

The cluster underneath: GKE charges $0.10 per cluster per hour, about $72 for a full month. Google’s $74.40 monthly credit covers one single-zone cluster’s fee per billing account. The machines are billed on top. So AX is free the way a free puppy is free.

The voice: At Google’s published rates, Gemini 3.8 Live costs about half a cent per minute of caller audio and about 1.8 cents per minute of the agent’s voice. Extended Thinking bills its thinking at the voice-output rate, so the same call costs more. In Artificial Analysis’s cost test, an hour of input audio ran $0.84 on base Live and $3.50 on Extended Thinking.

A worked example: a seven-minute call where the agent talks the entire time comes to about 16 cents at those per-minute rates. That’s before any thinking tokens and before your phone provider’s per-minute fees.

The server: $25.99 a month for the first two years on the KVM 8.

What they said vs. what the tests showed

Claim: Agent Substrate runs agents at 10 times the density of standard containers (Google, September 15, 2026).
The evidence: No independent benchmark exists yet. Solo.io’s June 12, 2026 walkthrough reported about 250 agent sessions running on 8 Kubernetes pods. That’s a third-party demo, not a controlled test. Treat 10x as vendor-reported.

Claim: Google’s docs call Gemini 3.8 Live the default for most low-latency voice agents (Google AI for Developers, September 2026).
The evidence: Artificial Analysis’s τ-Voice test checks whether a voice agent finishes multi-step tool tasks. On it:

  • Base Live scored 30.1%.

  • The previous Gemini 3.1 Flash Live scored 37.7%, so base Live came in below its own predecessor.

  • Extended Thinking scored 68.6%.

(Reported by DataCamp, September 17, 2026.)

Human listeners rank the two models the other way. In the blind Speech Agent Arena, Extended Thinking sits seven places behind base Live (reported by Heise Online, September 2026). So base Live sounds better, and Extended Thinking finishes more bookings. Test both on real calls.

Claim: 62% of calls to small service businesses go unanswered. The figure comes from a 411 Locals study of 85 small businesses (as reported by Automatdo, April 3, 2026).
The evidence: Invoca’s call data across home-service contractors shows a 27% missed rate, a 2024 to 2026 rolling average (as reported by Grovi, June 9, 2026). Both numbers come from companies that sell phone tools. Consider the figure disputed. Both sources agree that after-hours calls are where the misses pile up.

So what

Do this: Build the voice receptionist on one server, one app for all customers, with every event logged from day one.

Skip this: Kubernetes and AX for your first 25 customers.

Wait on this: AX in production. No stable release yet, and production support is still by application.

Steal this line: “The phone agent is employee number one. The business is the other nineteen.”

For most of history, a six-person roofing crew couldn’t afford a front office. Somebody’s spouse answered the phone between school pickups, and the calls that came in during a storm just rang out. That’s changing. The voice agent gets you in the door, and a sleeping AI back office is where the real business grows.

Build the receptionist first. Log the events. When your bill starts showing thousands of agents sitting around waiting, that’s your cue to call AX.

Talk soon.

Sources

What AX does, in plain English

Why AX stays out of a live phone call

The $26 setup for your first customers

The When-to-Switch Table

What it costs, and the license

What they said vs. what the tests showed

]]><![CDATA[Four AI Launches in 48 Hours, One Story Nobody Connected.]]>https://aisignal.veletica.com/p/four-ai-launches-in-48-hours-onehttps://aisignal.veletica.com/p/four-ai-launches-in-48-hours-oneWed, 23 Sep 2026 18:38:21 GMT

By the end of this, you’ll know which model to hand which job, and roughly what each finished job costs you.

The verdict first. On Sept. 22, OpenAI and Anthropic shipped cheaper models 90 minutes apart, one day after SpaceXAI shipped Grok 4.7. The same week, Google’s open-source agent runtime, AX, became the most discussed AI post on Hacker News. Most coverage is ranking who got smarter. The bigger change is the price of keeping smart AI working on a job for hours. On independent tests, GPT-6 Sol does the same benchmark work as its predecessor for about half the money. Claude Opus 5.5 scores higher and costs a bit more per job. For most people, Sol is the new default, and Opus 5.5 is the one you call when the answer has to be right.

How AI billing works, in plumber terms

Think about hiring a plumber. The hourly rate is right there on his website. The invoice is what makes you sit down. A $60-an-hour plumber who needs three hours costs you $180, while the $100 guy who's done in one hour costs you $100 and leaves before your coffee gets cold.

AI works the same way. Companies advertise a price per million tokens. A token is a chunk of text, roughly three-quarters of a word. That’s the hourly rate. What you pay for a finished job depends on how many tokens the model burns getting there, and agents burn a lot, because they read, try, check, and try again.

Then there’s cached input, which is text the model already read that it can reuse at a discount. Paying full price to re-read it every turn is like your plumber charging you to find your kitchen again every time he walks back from the van. Five trips means five kitchen-finding fees. By trip six you’re wondering if he’s billing you for the van too.

This week’s launches go after the invoice. OpenAI now discounts cached reads by 90%. Anthropic cut Opus 5.5’s cached reads 60%, to $0.20 per million tokens. And Lovable, a company that builds apps with AI, told Anthropic that Opus 5.5 finishes jobs in a third to half fewer steps. Fewer steps means a smaller invoice, even before anyone touches the hourly rate.

What you can hand an agent now that was too pricey in August

This part changes what a regular person or a five-person team can get done.

You can hand off an overnight job. A developer at the legal software company Clio gave Opus 5.5 an engineering task across six code repositories and let it run unattended. It stayed on the job for more than 18 hours and needed minimal rework, in his words on Anthropic’s launch page. An agent, for anyone new here, is AI that takes actions on its own. A chatbot tells you how to rename the files. An agent renames them.

You can keep a big pile of documents loaded without going broke. Say your agent works from 50,000 tokens of contracts, notes, or a product manual, and re-reads them every turn. On GPT-6 Sol, cached reads cost $0.20 per million tokens, so each re-read costs about 1 cent. At Sol’s full input rate, it would be about 10 cents. Over a few hundred turns, that gap buys lunch.

You can run research that gets fact-checked. In Anthropic’s own test, Opus 5.5 wrote company earnings reports from a copy of the web where the real numbers were hard to find. A grader failed any report with an invented figure or quote. Opus 5.5 passed 16 of 18 times, and Anthropic says Opus 5 and Fable 5.1 passed none. That’s the vendor grading its own homework, so treat it as a promising sign.

You can put AI on paperwork. In a Box demo SpaceXAI promoted, Grok 4.7 checked a $2 million insurance claim against the policy and caught $146,000 in duplicate invoices and missed deductions, the kind of thing a tired human skims past at 4 p.m. on a Friday. It's a demo, so expect less on your own files.

And you can send routine work to the cheap seats. GPT-6 Luna ran Artificial Analysis’s independent test tasks for about 7 cents each. At that price, you can hand Luna the grunt work, like sorting your inbox or tagging support tickets.

The cost-per-job table

Artificial Analysis, an independent testing firm, runs every model through the same set of tasks and publishes what each run costs. A higher score means smarter on their test. “Effort” is a dial for how long the model thinks before answering. More effort usually means better answers and a bigger invoice.

The AI Signal Receipt: what a finished job costs

Checked Sept. 22, 2026. Scores and per-job costs come from Artificial Analysis’s independent test set. This market reprices every few weeks, so re-check before you budget.

If you’re spending less than about 55 cents a job, Sol scores higher than any Opus 5.5 setting at the same price. If you can spend more, Opus 5.5 at its default setting beats Sol at its maximum, and it costs about 26% more per job ($1.34 vs. $1.06).

On licensing: every model here is paid and closed, billed per token through each company’s API. The chat apps are a different deal. GPT-6 Sol and Luna are in ChatGPT’s paid plans and Codex, and Free and Go users get Luna in the desktop app. Opus 5.5 is on Claude’s paid plans. Google’s AX is free and open source under the Apache 2.0 license, which lets you use and modify it commercially. Free software still needs servers, though. AX expects a Kubernetes cluster (a system for running lots of software containers across many computers) plus a container registry, and it still calls paid models to do the thinking. A free runner wrapped around a paid brain is still a paid setup.

Where the cheap price tag lies

Claim: OpenAI says Sol and Luna are 50% cheaper (Sept. 22). Receipt: OpenAI measured that against GPT-5.6’s promotional pricing, per its own announcement. Artificial Analysis measured about 50% lower cost per task for Sol and about 60% for Luna (Sept. 22), and found their intelligence scores roughly level with GPT-5.6. So they’re cheaper, and about as smart as before.

Claim: Anthropic says Opus 5.5 costs about 40% less to run than Opus 5 on typical work (Sept. 22). Receipt: Digital Applied points out that the saving compares each model at its default effort, which is medium on Opus 5.5 and high on Opus 5 (Sept. 22). The saving is real for anyone using defaults.

Claim: SpaceXAI says Grok 4.7 is twice as fast at half the price of comparable models (Sept. 21). Receipt: Grok 4.7’s output tokens cost less than Sol’s ($6 vs. $10 per million), and Artificial Analysis still measured $2.73 per test task, more than double Sol at max (Sept. 22). The reason is chattiness. Artificial Analysis counted roughly 81,000 output tokens per task at its highest setting, against 36,000 for Grok 4.6 (reported Sept. 22). It has a lower hourly rate than Sol and a much bigger invoice.

Claim: Google’s AX repository says it can run billions of autonomous agent workloads. Receipt: there’s no published benchmark of that yet. The project’s own demo showed about 250 agent sessions across 8 machines, and its README warns of major breaking changes before a stable release (reported Sept. 22).

Claim: the 18-hour unattended run. Receipt: that’s one customer quote on Anthropic’s launch page (Sept. 22), with no neutral test behind it yet. It’s encouraging, and it isn’t proven.

The two-model setup that sends routine work to Luna

The setup I’d copy looks like this. Picture a weekly competitor price report. Luna visits the pages and pulls the prices into a list. Sol writes the report. Opus 5.5 gets called once, at the end, to check every number against the source. The cheap model goes first, and the expensive one goes last, only when it’s needed.

Put the stuff that never changes, your instructions and reference documents, at the start of every request, so the cache can reuse it. OpenAI now lets you change the effort dial or switch tools mid-job without losing that cache, and GitHub told OpenAI these caching changes more than halved the text Copilot had to process from scratch, across billions of requests.

For two years the question at every launch was how smart the new model is. The question on your desk now is how many of these you can afford to leave running, and which job you’d hand over first. Agents still fail. They still need someone reading the output in the morning. But the price of trying just dropped, and I don’t think it’s going back up.

So what

Do this: Pick one job that eats your Friday afternoon every week. Run it through GPT-6 Sol at high effort, then check your bill to see what that one finished job cost. That number beats any chart.

Skip this: Choosing a model by its price per million tokens alone. Grok 4.7 charges less than Sol per output token and costs the most per test task in the table.

Wait on this: Google’s AX, unless you already run Kubernetes. The README itself warns of major breaking changes.

Steal this line: “Don’t ask what the tokens cost. Ask what the finished job costs.”

Sources:

The verdict

How AI billing works, in plumber terms

What you can hand an agent now that was too pricey in August

The cost-per-job table

Where the cheap price tag lies

The two-model setup that sends routine work to Luna

Deep dive and weekly updates: subscribe for free at aisignal.veletica.com.

]]><![CDATA[The Claude Opus 5.5 Cost Guide: How to Hand Off Bigger Jobs for Less.]]>https://aisignal.veletica.com/p/the-claude-opus-55-cost-guide-howhttps://aisignal.veletica.com/p/the-claude-opus-55-cost-guide-howTue, 22 Sep 2026 19:30:21 GMT

Medium vs. Max Effort: Which One You Actually Need

By the end of this, you’ll know which jobs to hand Claude Opus 5.5 this week, which effort setting to pick, and roughly what a finished job costs you.

Anthropic released Opus 5.5 on September 22, 2026. My verdict: this is the first AI launch where the bill is more interesting than the brains. Anthropic says it costs 40% less to run than Opus 5 on typical workloads at default settings. The same day, the independent testing firm Artificial Analysis put it at the top of its intelligence rankings. Put those two together and you get a model you can afford to leave running for hours.

And people are leaving it running. A developer at the legal software company Clio handed it a job spread across six code projects, walked away, and let it work overnight. It stayed on task for more than 18 hours.

Meanwhile, I can’t stay on task for 18 minutes if somebody drops a meme in the family group chat.

The idea that makes this click: your “unit of work”

Every AI tool has a unit of work. That’s the biggest chunk of a job you can hand it and then go do something else.

Two years ago, the unit was “write me this email.” Last year it grew to “build me this feature.” With Opus 5.5, it looks more like “move this whole system to the new version and tell me when it’s done.”

So the useful question has changed. You used to ask how smart the AI is. Now you ask how big a job it can finish without you, and what the finished job costs.

Two quick definitions will make the rest easier to read. Tokens are the small chunks of text an AI reads and writes, a bit less than a word each, and you pay per token. Effort is how hard the model thinks before it answers. Opus 5.5 has five settings: low, medium, high, xhigh, and max. Medium is the default.

That’s why one boring line on Anthropic’s pricing page matters so much. When an agent works for hours, it keeps rereading the same material, like your project’s code. Those rereads are called cache reads, and Anthropic says they make up the majority of the cost of agent and coding work. Opus 5.5 cut the price of cache reads by 60%.

The cost-per-job table

This is the table I’d pin to the fridge. Every row is reported by Anthropic or one of its early-access customers, so treat each one as a best case.

Look at the Quantium row for a second. Thirty-eight prompts dropped to eleven. That’s 27 fewer times somebody had to stop what they were doing and babysit the AI. For most of us, that’s the real cost nobody puts on an invoice.

Jobs you can hand off this week

This is the practical part. Think about the way you used AI last month, then make the ask one size bigger.

  • Last month: “Summarize this report.” Now: “Research these three companies, build a comparison spreadsheet, and turn it into a five-slide deck.” That’s the shape of Anthropic’s merger test.

  • Last month: “Fix this one function.” Now: “Move the whole app to the new version and list what broke.” One early tester finished a 680,000-line code migration in less than a day. Anthropic says that work would have taken an engineering team weeks.

  • Last month: “Check these 20 files.” Now: “Audit the whole codebase and fix what you find.”

  • Last month: “Find me some numbers on this.” Now: “Pull last quarter’s figures, check every one against a source, and flag anything you couldn’t verify.” In one Anthropic test with a hard-to-find earnings release, 16 of its 18 reports passed a check where a single invented figure meant failure. Fable 5.1 and Opus 5 didn’t pass once.

You don’t have to write code for any of this to matter. The research and deck jobs work the same way for a small business owner, a teacher building a unit plan, or anyone with a pile of documents and a deadline.

Three habits make bigger handoffs go better:

  • Write down what “done” means. “Every figure has a source.” “All tests pass.” A finish line keeps a long run from wandering.

  • Say what it can’t touch. Name the files, accounts, or folders that are off limits. Anthropic says Opus 5.5 is much less likely than recent models to act outside the boundaries you give it, and clear boundaries give it something to respect.

  • Check at milestones. A long run is still worth a peek at the halfway point.

Medium vs. max: when to pay for the scratch paper

This comparison makes effort settings click. Medium does the job. Max does the job and keeps all its scratch paper, and you pay for every page.

The numbers back that up. Artificial Analysis found that at max effort, Opus 5.5 writes about 119,000 output tokens per task on its test suite. Opus 5 at max wrote about 73,000. That’s 1.6 times as much, and it’s why the independent testers found Opus 5.5 at max costs about the same per task as Opus 5.

So the famous 40% savings mostly comes from the default medium setting. And medium is strong. Anthropic says medium Opus 5.5 beat Opus 5 at max on the coding test Terminal-Bench 4.0, for about a fifth of the cost. Deloitte reported that even the lowest setting caught 72% of known bugs in its code reviews, where Opus 5 at high caught 56%.

Use medium when the job is clear and you can check the result. Use max when a wrong answer is expensive, like a financial model going in front of a board.

What it costs and how it’s licensed. Opus 5.5 is a paid, proprietary model. It isn’t open source, and there are no weights to download. At Anthropic’s published rates, the API costs $4 per million input tokens and $20 per million output tokens, both 20% lower than Opus 5. Cache reads are $0.20 per million, and cache writes are $5. A fast mode runs up to 2.5 times quicker at $8 input and $40 output per million. If you use the Claude app on a Pro, Max, Team, or seat-based Enterprise plan, Anthropic raised your five-hour usage limits and added a rate-limit reset you can save for later. It’s also on Amazon Web Services, Google Cloud, and Microsoft Azure.

Two limits are worth knowing. Thinking can’t be switched off on this model, so every answer includes some reasoning you pay for. And safety filters send most cybersecurity tasks to the older Opus 4.8, while flagged biology requests go to Opus 5.

Claims vs. Receipts

Claim (Anthropic, Sept 22, 2026): 40% cheaper to run than Opus 5 on typical workloads.
Receipt (Artificial Analysis, Sept 22, 2026): At max effort, it lands about level with Opus 5 on cost per task, because it writes 1.6 times the output tokens. Both can be true. The saving depends on your effort setting.

Claim (Anthropic, Sept 22, 2026): Performs at the level of Claude Fable 5.1 on most work.
Receipt (Artificial Analysis, Sept 22, 2026): At max effort it scores 58 on the Intelligence Index, ahead of Fable 5.1 and GPT-6 Astra at 53 each. At high effort, it scores 54.

Claim (Anthropic and partners, Sept 22, 2026): A 680,000-line migration in under a day, and an 18-hour unattended run.
Receipt: These haven’t been independently reproduced yet. They’re real reports from named companies, and they show what’s possible on a good day.

Anthropic published one more receipt about itself. The company says Opus 5.5 often seems to suspect it’s being tested, which makes it harder to predict how it acts in the real world. I respect a launch post that admits that.

So what

Do this: Pick one job you usually split into ten prompts. Hand it over whole, at medium effort, with a written definition of done.

Skip this: Max effort as your everyday default. You’ll pay for scratch paper you never read.

Wait on this: Sonnet 5.5 and Haiku 5.5, which Anthropic says arrive in the coming weeks. If your jobs are small and high volume, a cheaper model may fit better.

Steal this line: “Stop asking how smart it is. Ask how big a job it can finish without you, and what the finished job costs.”

For most of us, the limit on what we could build was never a lack of ideas. It was hours in the day. This week, that limit moved a little. Go hand something big to your new overnight coworker, and please, unlike me, get some sleep while it works.

Talk soon,
Oscar

Receipts

The launch and the verdict

The idea that makes this click

The cost-per-job table

Jobs you can hand off this week

Medium vs. max

Claims vs. Receipts

So what

]]><![CDATA[The AI Agent Hands Field Guide: Know What Your AI Can Touch Before You Pay for It.]]>https://aisignal.veletica.com/p/the-ai-agent-hands-field-guide-knowhttps://aisignal.veletica.com/p/the-ai-agent-hands-field-guide-knowMon, 21 Sep 2026 04:51:09 GMT

Sunday AI Signal: one pick worth your attention, what the week’s coverage oversold, and what to do about it on Monday.

By the end of this, you’ll have one question that tells you whether any AI product is worth your money.

For two years I’ve graded chatbots like a talent show judge. I wanted to know which one was smartest and which one wrote the best emails. Meanwhile, the smartest one in the room couldn’t turn off my kitchen light. It could write a sonnet about my kitchen light. It was a lovely sonnet, and the light stayed on.

That changed between September 15 and 18. Google, Meta, and Vercel each shipped a way for AI assistants to press real buttons inside real products. So here’s the verdict up front: stop asking which AI is smartest and start asking what it can reach. One of these is usable by a regular person this week, and it sits behind a $20-a-month plan. The other two matter mostly if you build things.

Quick definition, because the name helps nobody. MCP stands for Model Context Protocol. Think of it as a standard plug. A company builds the plug once, and any AI assistant that fits it can use that company’s product. An assistant that acts for you this way gets called an agent.

Examples that help me tell the two apart: a chatbot is the friend who gives great directions. An agent is the friend who has your car keys.

What your agent can reach now, and what stays locked

Google Home (September 16). You can ask your house what happened while you were out, and the answer comes from your devices’ real history. You can say “turn off all the outside lights,” and they go off. Google’s setup guide lists five tools the agent gets: find your homes, list your devices, check their status, read their history, and run actions. Google says early testers asked what the kids did after school and got a summary pulled from every camera. Another counted a week of laundry loads. I don’t want that number for my house.

What stays locked: unlocking doors is a blocked action. Creating automations isn’t supported yet, and Google lists slow responses as a known issue.

WhatsApp Business (September 15). You can set up WhatsApp for your shop by asking. Tell a coding assistant such as Claude, Cursor, Codex, or ChatGPT what you need, and it creates the business account, adds and registers your phone number, drafts your message templates, points customer replies at the right system, and sends a test. Meta says that job used to mean bouncing between four places: its Developer Console, Business Manager, the technical manual, and your code editor. You still type in the verification code Meta texts you, and Meta says anything that changes the account needs a signed-in person.

Vercel (September 18). This one is for people who build websites. WebMCP is the same plug, built into a web page, so an assistant living in your browser can use the site’s tools while you’re logged in. Vercel says a site that already has MCP tools can switch them on with a single script tag, which is one line of code on the page, plus a short list of which tools to allow. Nobody builds a second login system. It’s labeled experimental; WebMCP is still a proposed web standard, and the developers behind the change note that browser support is limited.

What it costs to get in

Google Home MCP: paid. Access requires Google Home Premium Advanced, which TechCrunch reported at $20 a month on September 16. You also need a Google Cloud project and an AI assistant that supports MCP, which for most people is another subscription.

WhatsApp Business Tools MCP: free to use. Meta hasn’t announced a charge for the tool. WhatsApp’s normal fees for paid business messages still apply once you go live.

Vercel’s WebMCP support: free and open source, experimental.

None of the three bills you per action. What you pay for is the subscription that gets you through the door, plus setup time. Google’s instructions have you create a cloud project, switch on the Home API, and generate login credentials. That reads like a weekend project for a tinkerer and a hard no for my mother.

Claims vs. Receipts

The week’s other big story was Jev, from a startup called TypeSafe AI, launched September 15. Jev doesn’t chat. You hand it messy information and a question with fixed answers, and it hands your software numbers back, such as fraud 0.96 and legitimate 0.04. A chatbot writes you a paragraph. Jev hands your software a number. Use a chatbot when you need an explanation. Use something like Jev when software has to make the same small call ten thousand times a day.

TypeSafe prices it at $0.042 per million input tokens with output free. A token is a chunk of text about three-quarters of a word long, so that’s roughly four cents to read 750,000 words. It’s a paid online service with a waitlist, and MarkTechPost reported on September 19 that TypeSafe hasn’t released a version you can download or run yourself. Tiny prices still add up: TypeSafe’s own demo of Jev playing the video game Doom made 10 decisions a second and cost about $7 an hour, by the company’s count.

Claim: “193.6x faster, 444.6x cheaper.” (TypeSafe, September 15) Receipt: the same TypeSafe post says it expects those figures to sit at the higher end of real-world gains. It also says TypeSafe’s own team built the four test jobs, and that Jev was graded against the averaged answers of two large chatbots instead of an answer key. Independent numbers are smaller. TechCrunch reported on September 18 that a Vercel engineer saw a safety check run 5 to 18 times faster after switching to Jev. In the same article, another developer found Jev 10 to 20 times cheaper than Gemini at sorting business email, with Gemini slightly more accurate. One independent GitHub test dated September 18 measured 0.43 seconds for Jev against 0.92 seconds for a small, cheap chatbot model.

Claim: Jev “can’t hallucinate.” (TypeSafe, September 15) Receipt: TypeSafe’s own footnote says its 0% figure “is not empirical.” The guarantee is that Jev can only pick from the answers you gave it. It can still pick the wrong one with a straight face. An independent evaluation called PriorBench, dated September 20, scored Jev at 95.9% on a 400-item test. The same evaluation watched it file a cake recipe as a “technical issue” at 0.94 confidence, and saw 0 of 30 off-topic messages get flagged when the answer list had no “none of these” option. We’ve all worked with this guy. He has never said “I don’t know” in a meeting. One caution on the receipt itself: PriorBench is a single anonymous author’s one-evening test, run through a middleman service.

Claim: AI agents can now control your locks. (A DEV Community write-up from this past weekend listed lights, locks, and thermostats.) Receipt: Google’s setup guide, last updated September 15, says unlocking doors is a blocked action and that the service enforces rate limits. Google also suggests telling your household before you connect an agent.

Skip / Wait / Steal

Do this: If you already pay for Home Premium Advanced and you’re comfortable following a setup guide, check whether access has reached your account. Google says it’s rolling out over the coming weeks.

Skip this: Upgrading to the $20 plan only for this. Slow responses are a listed known issue, and automations aren’t there yet.

Wait on this: Jev for any decision where a wrong answer costs money. The independent tests are days old. If you build software and try it anyway, PriorBench’s most useful tip is to always give it a “none of these” option.

Steal this line: “I stopped asking which AI is smartest. I ask what it can reach.”

The web spent decades building doors for software. Last week three big companies started building doors for AI assistants, in four days, without coordinating a press release.

Your one next step: open the AI tool you pay the most for and find its list of connections, sometimes called connectors or integrations. That list is what you’re really paying for.

Talk soon.

Sources:

What your agent can reach now, and what stays locked

What it costs to get in

  • Cost figures in this section come from the TechCrunch Google Home entry and the Google setup guide above.

Claims vs. Receipts

]]><![CDATA[The Sponsored Agents Decision Guide: Figure Out What Your Website Is For Before Your Customers Stop Visiting It.]]>https://aisignal.veletica.com/p/the-sponsored-agents-decision-guidehttps://aisignal.veletica.com/p/the-sponsored-agents-decision-guideThu, 17 Sep 2026 02:32:25 GMT

On Tuesday, Google shipped a voice model that keeps talking while it works. On Wednesday, Anthropic erased the line between chatting with Claude and handing Claude a job. The same day, OpenAI started testing ads inside ChatGPT that you can talk to.

That is three companies in two days, all moving the same way. The software is learning to skip the page and go straight to the person.

My verdict: if you sell anything online, the next twelve months are about deciding what your website is actually for. You do not need to redesign it yet. You need to decide what job it does once an agent can answer the first ten questions a customer used to bring to your homepage. The Sponsored Agents pilot is small, it is US only, and it covers a handful of advertisers, so nothing breaks tomorrow. But the shape of it is the whole story, and I would rather you see the shape now than read about it in somebody’s earnings call next spring.

The showroom and the parking lot

For thirty years the internet worked like a car dealership. You searched, you clicked, you walked into the showroom, and eventually a person found you. Every marketing dollar on earth was spent getting you through that door.

A Sponsored Agent is the salesperson walking out to the parking lot. You see an ad in ChatGPT, you tap it, and a clearly labeled chat opens with an agent that speaks for the business. You explain what you need, it asks follow-up questions, and when you are ready it hands you a link. OpenAI says that chat stays separate from ChatGPT’s own answer and separate from the conversation you started. As of September 16, 2026, the test is limited to select advertisers in the United States.

Two definitions, because I promised my mom this newsletter would stay readable. An agent is software that can ask you questions and take a next step for you, instead of only answering. A website is a brochure on a table. Use the website when someone already knows what they want. Use an agent when they only know what is wrong.

I spent years assuming the homepage was the front door. It turns out the front door can walk.

1. What you can do this week with each launch

Talk to an ad. If you are a US ChatGPT user on a tier that shows ads, you may already see an Angi agent appear in a home repair conversation or a Wayfair agent in a furniture one. You can describe the leaky faucet, answer a few questions, and end up in Angi’s service request flow without ever loading angi.com. Angi says so in its own press release from September 16. Wayfair told Contentgrip it is participating at limited scale.

Build a campaign by talking. Advertisers can now create, update, and analyze ChatGPT ad campaigns with plain prompts in ChatGPT Work through the Ads Manager plugin. You paste a website or a brief, ask for a campaign, and get suggested copy and images pulled from your own landing page. You still approve every line before it runs.

Run it where you already work. HubSpot, a customer database most sales teams live in, and Shopify, the online store builder, both got ChatGPT Ads apps on September 16. The HubSpot listing already shows more than 400 installs and adds tracking tags so an ad click maps to a real contact. The Shopify app is US only until September 23, when it reaches the other markets where ChatGPT Ads run.

Hand Claude the whole job. Anthropic merged Cowork, its longer running work mode, into regular Claude chat on Pro and Max plans, rolling out over the coming weeks. You type “what moved in the pipeline last week, write it the way we always do, and put the highlights in five slides,” then close your laptop. Claude decides whether that is a chat, a research run, or a document. Claude Docs and Slides shipped in beta the same day. By default Claude asks before it acts. You can switch that so it only checks in when something needs a closer look.

Keep talking while Gemini works. Gemini 3.8 Live runs tool and API calls in the background while the conversation continues, and it switches between 97 languages mid-sentence. The Extended Thinking version narrates its progress on multi-step jobs, so you hear “let me check that” while it books the table or builds the page. Since September 15, Extended Thinking is free in Gemini Live, and Live is free in Search Live.

Here is the Pages vs. Agents table so you can hold all three in one glance.

2. The Angi handoff, step by step

  1. You are in ChatGPT asking why your water heater sounds like a haunted house.

  2. A labeled Angi ad appears. You choose to open the chat with the Angi agent.

  3. The agent asks about the unit’s age, the noise, and whether there is water on the floor.

  4. When the job needs a professional, it moves you into Angi’s service request flow to connect with local pros.

  5. Angi gets a qualified lead. You never opened a browser tab.

Angi already had an opt-in ChatGPT app since March. The difference is who makes the first move. You add the app yourself, while the sponsored agent shows up on its own inside a relevant conversation. That one difference is the entire business model.

I have spent an embarrassing share of my adult life fixing things at 11 pm with a headlamp and a YouTube video from a man named Gary. The version of me who could describe the noise out loud and get handed a plumber is the version who still has his security deposit.

3. What it costs: claims vs. receipts

ChatGPT Ads run on an auction. CPC means you pay per click. CPM means you pay per thousand views. Both are available.

Vendor claim (OpenAI, per its Help Center as quoted by several ad guides): a recommended starting bid of $3 to $5 per click, and a $60 default maximum for per-thousand-view bidding.

Independent figures (June through August 2026): guides from WebFX, Lapis, and Context Hints report clearing prices of roughly $25 to $60 per thousand views. The minimum spend went from a reported $200,000 at the February launch, to $50,000 in April, to removed on May 5, 2026 when the self-serve Ads Manager opened. One August guide (SEO Discovery) says a $25 daily minimum applies in the US. Those two claims conflict. I could not resolve it from OpenAI’s own pages, so check ads.openai.com before you budget a dollar.

Sponsored Agents specifically: no published price. According to The Deep Dive’s read of OpenAI’s Ad Tools Terms, the advertiser is responsible for the agent’s content, configuration, actions, and output. In plain English, if your agent promises a discount you do not honor, that is your problem, not OpenAI’s.

Claude: $20 a month on Pro, $100 on Max 5x, $200 on Max 20x. Docs, Slides, and Design are in beta on paid plans. Team and Free plans follow later, and Enterprise admins get 30 days notice before anything changes.

Gemini: Extended Thinking is free in Gemini Live and Live is free in Search Live. Developers pay through the Gemini API. Google’s numbers, published September 15, are vendor reported: 68.6% on the τ-Voice task test, 35.1% on Sierra’s banking version, and an 82.6 on Artificial Analysis’ Speech to Speech Quality Index. Artificial Analysis is an independent tracker, but Google picked which scores to put in the blog post. I have not seen a third party rerun them yet.

4. Five questions to ask about your own website

  1. If an agent answers the first ten questions, what is left for the page to do?

  2. What is your agent allowed to say, promise, or discount? You own its mistakes.

  3. When does it hand the customer to a human, and does a human know it happened?

  4. How do you count a conversation? A click is a number. A ten-minute chat that ends in a lead is a different number, and your dashboard does not have a column for it yet.

  5. Are you still optimizing for humans clicking, or for agents finding you and choosing you?

I have spent two decades learning systems right before somebody replaced them, so I know the feeling. Now the thing a web page competes with is a chatbot in a parking lot, and the chatbot is nicer than I am.

So what

Do this. Read your homepage out loud as if you were an agent trying to answer a stranger’s question in one reply. Every sentence it cannot use is a sentence a Sponsored Agent will skip. If you already run ads and you are on Shopify or HubSpot, a small test at OpenAI’s $3 to $5 starting bid tells you more than another month of reading about it.

Skip this. Building your own Sponsored Agent right now. It is invite only, US only, and the terms put every word it says on you.

Wait on this. Google’s voice benchmarks. They are strong, they are vendor picked, and nobody independent has rerun them as of September 16.

Steal this line. “Use the website when someone knows what they want. Use an agent when they only know what is wrong.”

Thanks for reading this far. If a plumber shows up at your house because a chatbot sent him, tell him I said hi.

Receipts

The showroom and the parking lot

  • OpenAI: Reimagining advertising with AI (September 16, 2026). Backs the Sponsored Agents test, the “clearly labeled” and “separate conversation” claims, and the select US advertisers limit.

1. What you can do this week with each launch

3. What it costs: claims vs. receipts

]]><![CDATA[Gemini 3.8 Live: Google Just Taught Its Voice AI to Walk and Chew Gum. Here's How to Use It.]]>https://aisignal.veletica.com/p/gemini-38-live-google-just-taughthttps://aisignal.veletica.com/p/gemini-38-live-google-just-taughtWed, 16 Sep 2026 19:07:25 GMT

A plain-English setup guide to Gemini 3.8 Live and 3.8 Live Extended Thinking, with the real cost per hour and the two things the launch coverage got wrong.

You know that moment when you ask a voice assistant something slightly complicated and it just... goes quiet? You stare at the little pulsing circle. You say “hello?” like it’s a bad phone connection. Then it comes back with an answer to a question you’ve already forgotten.

Google says that moment is over. On September 15, it released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two voice models that can run tools and reason in the background while they keep talking to you.

Here’s my verdict, so you can stop reading if you’re busy. If you just talk to Gemini in the app, you don’t have to do anything. It’s rolling out to you now. If you build things, the rule is simple: Live talks while it fetches; Extended Thinking talks while it thinks. Start with plain 3.8 Live and only move up when your agent has to work through a hard problem mid-sentence. Budget about $1.38 for an hour of nonstop conversation at Google’s published rates, and expect more than that the moment the thinking version starts thinking hard.

Now the part that actually matters: what this lets you do that you couldn’t do last week.

The waiter who never leaves your table

Every voice assistant you’ve used until now has been a waiter who walks to the kitchen every single time you ask a question. Want the specials? He leaves. Is the fish fresh? He leaves again. You make small talk with the salt shaker.

The new model is a waiter who yells your order back to the kitchen while still standing at your table, telling you about the fish. The kitchen is the tools, the calendar lookups, the API calls, the deep reasoning. It’s all running in the background. The waiter never stops talking.

That’s the whole story. Google’s engineers call it asynchronous function calling. I call it the end of hold music.

Five things you can do now that you couldn’t do before

  1. Ask for three things at once and keep the conversation going. Say “check the weather, pull up Thursday’s calendar, and while you’re at it, explain why my sourdough keeps collapsing,” and the agent starts the lookups and starts talking about your sad bread. When the tools finish, it folds the results back in. Before this, every one of those requests was a full stop.

  2. Point your camera at something and talk about it while the agent works. Both models process visual input in near real time. Google’s launch demos show the model narrating a chess game and walking someone through a software onboarding flow using what’s on screen. If you’ve ever read an error message out loud to a support line, you know why this matters.

  3. Switch languages mid-sentence and have it follow you. The models detect and switch among 97 supported languages during a conversation, automatically. Anyone who runs bilingual meetings knows the pain of assistants that lock onto one language and refuse to let go.

  4. Hear progress instead of silence. This is the Extended Thinking specialty. It uses early verbal cues like “Let me check that...” and then narrates its progress while it works through multi-step tasks in the background. You still wait for the hard stuff. You just don’t wait in the dark.

  5. Build a voice agent with one model instead of three. Most voice agents have been a relay race: one model to transcribe, one to think, one to talk back. Every handoff added delay. These are native speech-to-speech models, so the whole thing is one runner.

Which Gemini is which: the one table you need

Quick clarification before the table, because the names are a mess. There is no “Gemini 3.5 Live” conversation model. The 3.5 audio models are specialists: Gemini 3.5 Transcribe turns speech into text, and Gemini 3.5 Live Translate turns speech in one language into speech in another. The actual family tree for talking to an agent is 3.1 Flash Live, then 3.8 Live, then 3.8 Live Extended Thinking. Think of Live as the receptionist who can look things up while chatting, and Extended Thinking as the analyst who narrates while she works the problem.

The short read: if you’re on 3.1, Google is telling you to move. If you’re starting fresh, take 3.8 Live and don’t touch Extended Thinking until your agent needs to think while it talks.

How to use it, path one: just talk

If you’re not a developer, your entire to-do list is “open the Gemini app and tap Live.” Extended Thinking started rolling out to Gemini Live on September 15. It’s also behind the new Gmail Live (talk to your inbox), Docs Live (draft and edit by voice), and Keep Live (take notes by talking). Search Live gets it too.

Rollouts are gradual. If your app still hangs when you ask something complicated, you haven’t gotten it yet. Give it a few days before you assume it’s broken. It’s the rollout.

How to use it, path two: build something

Step 1: Open Google AI Studio. Go to the Live playground and pick a model. There’s a free tier, and the catch on the free tier is that Google uses your content to improve its products. Switch to paid, and it doesn’t.

Step 2: Pick the right model. The model string is gemini-3.8-live for the default and gemini-3.8-live-extended-thinking for the heavy version. Google’s own docs say 3.8 Live is the default for most voice agents. Live talks while it fetches. Extended Thinking talks while it thinks. If your agent’s hardest job is looking things up, stay on Live. If its hardest job is reasoning, go Extended.

Step 3: Learn the four rules that will break your code. Nobody’s tutorial mentions these until you’ve lost an afternoon.

  • On Extended Thinking, turnComplete: true no longer means the model is done. It may still be reasoning or running tools. Watch the interaction_status field instead: IN_PROGRESS means more is coming, IDLE means it’s actually waiting on you.

  • Extended Thinking only allows asynchronous function calls (behavior: NON_BLOCKING). Blocking mode returns a hard error. On plain 3.8 Live, async is the default, but you can still set BLOCKING for older code.

  • thinking_level (low, medium, high) only works on Extended Thinking. Plain 3.8 Live rejects it, so remove it from your setup if you’re migrating. MINIMAL isn’t supported on either.

  • Proactive audio is permanently on. Affective dialogue is gone from the API. And video frames are sent to the model by default, so only send frames you actually need, because every frame costs money.

Step 4: If you’re migrating from gemini-3.1-flash-live-preview, swap the model string, strip out any thinking_level and enable_affective_dialog settings, and turn on output audio transcription if you need a text transcript. Audio is the supported response mode.

What it costs and who owns it

These are hosted, closed-weight models. No self-hosting, no open license. You rent them through the Gemini API, AI Studio, or Gemini Enterprise.

At Google’s published rates, audio input is $0.005 per minute and audio output is $0.018 per minute. Image and video input is $0.002 per minute. Text runs $0.75 per million tokens in and $4.50 per million out, and that output price includes thinking tokens.

So an hour where you talk the whole time and it talks the whole time is $0.30 plus $1.08, which is $1.38. That’s a ceiling for plain conversation, not a bill for Extended Thinking. Artificial Analysis, as reported by OfficeChai on September 15, measured 3.8 Live at about $0.84 per hour of input audio on its Big Bench Audio test and Extended Thinking at about $3.50. The gap is the thinking tokens. The more it reasons, the more you pay, and you can’t see the meter while it’s running.

Two more line items that sneak up on people: Google Search grounding gives you 5,000 free requests a month shared across all Gemini 3.x models, then $14 per 1,000. And those default video frames add up if you leave the camera on.

Claims vs. Receipts

Claim: Google didn’t publish per-minute pricing for these models. Shattered.io wrote on September 16 that Google hasn’t done so, leaving cost as “the biggest open question.” Google’s pricing page, updated September 15, lists $0.005 and $0.018 per minute in plain text. The receipt wins.

Claim: Google beat OpenAI on voice quality. The 82.6 score on Artificial Analysis’ Speech to Speech Quality Index belongs to Extended Thinking, not the standard model. OfficeChai, citing Artificial Analysis on September 15, puts standard 3.8 Live at 76.0. OrcaRouter argued the same day that, standard tier against standard tier, OpenAI’s GPT-Live-1 wins and “it is not close.” OpenAI’s own index number is disputed in the coverage: OfficeChai cites 81.5 for GPT-Live-1 Astra (Medium), while AutoInterviewAI cites 79.4. Treat any single OpenAI figure with suspicion until Artificial Analysis’ own page settles it.

Claim: Gemini is 13 times cheaper than OpenAI. AutoInterviewAI published that multiple times on September 16. Artificial Analysis’ measured cost per hour, via OfficeChai, shows 3.8 Live at $0.84 versus GPT-Live-1 Astra at $5.83, which is about 7 times. Extended Thinking at $3.50 is under 2 times cheaper. The multiple depends entirely on which tier you compare and how much reasoning you trigger.

So What

Do this: If you build voice anything, open AI Studio this week and run gemini-3.8-live with one async tool call. Watch what the conversation feels like when the agent keeps talking during the lookup. That’s the feature. Everything else is a benchmark.

Skip this: Don’t start on Extended Thinking. It costs more, it changes how “done” works in your code, and most agents don’t need it on day one.

Wait on this: Any head-to-head with OpenAI. The numbers in circulation don’t agree with each other yet. Give Artificial Analysis two weeks.

Steal this line: “Voice stopped being an input and became a place where work happens.”

For twenty years, talking to a computer meant talking, stopping, and waiting. Google just took the stopping out. Whether you build agents or just want your inbox to answer back, that changes what a conversation with a machine can be.

Receipts

The release

Five things you can do now

How to use it

Which Gemini is which

What it costs

Claims vs. Receipts

]]><![CDATA[Salesforce Just Built Its Own Brain. Your Company Can Too.]]>https://aisignal.veletica.com/p/salesforce-just-built-its-own-brainhttps://aisignal.veletica.com/p/salesforce-just-built-its-own-brainWed, 16 Sep 2026 12:51:15 GMT

Salesforce built its own reasoning model on open weights. A startup raised $40 million to inspect AI agents before they touch your systems. Both landed on the same day, and together they are a blueprint you can copy small.

Two announcements dropped on Tuesday, September 15, 2026. Read separately, they’re a product launch and a funding round. Read together, they’re a preview of what enterprise AI looks like once it grows up.

My verdict: Salesforce’s new Koa model proves that a company with deep expertise in one job can now build its own specialized reasoning model instead of renting every ounce of intelligence from OpenAI, Anthropic, or Google. AIUC’s $40 million raise proves the next question buyers will ask is not “is this agent smart enough” but “who checked that it won’t leak my data.” Own the routine thinking, rent the hard thinking, and get both inspected before they go live. That’s the play, and a small version of it fits in a weekend.

The landlord and the garage

For two years, most companies have been renting their AI the way you rent an apartment. Every request goes to the landlord, the landlord sets the price, and you never own anything. It’s a nice building, though.

Koa is Salesforce deciding to build a garage instead of a mansion, with one very specific car parked inside.

Salesforce took NVIDIA’s open-weight Nemotron 3 Super model and post-trained it on a synthetic dataset modeled on 27 years of CRM workflows: qualifying leads, routing service cases, scheduling follow-ups. No customer data went in. Salesforce says the scenarios simulate workflows across more than 14 industries, each one pairing a persona with a task and mapping the exact sequence of tool calls an agent needs to finish it.

Then Salesforce kept the weights. Post-training and inference run inside its own infrastructure, so customer data never leaves its trust boundary. Jayesh Govindarajan, Salesforce’s EVP of AI, told TechCrunch the company had always leaned on frontier labs for reasoning, “until now,” and that the missing piece before Nemotron was a US-made, open, state-of-the-art base model with clear data provenance.

I’ll be honest about my reaction. My first thought was “great, Salesforce has a model, I will never touch it.” My second thought was the one that matters: if the recipe is open weights plus your own expertise plus synthetic data, the recipe isn’t exclusive to Salesforce. It’s just expensive right now, and expensive things get cheap.

The part Salesforce isn’t saying loudly

Salesforce is not breaking up with the frontier labs. Three weeks before Koa, on August 26, it announced ClaudeForce with Anthropic, which lets companies use Claude as the interface while their data stays inside Salesforce. TechCrunch reports that Agentforce’s AI gateway routes each request to whichever model fits: Koa for CRM reasoning, a frontier model when the job is harder.

The picture is a stack. Open weights become specialized models. Specialized models sit behind a router. The router calls a frontier model on demand. Salesforce just published the wiring diagram, and I think it’s what most serious companies will end up running.

Meanwhile, someone showed up with a clipboard

The same day, AIUC (the Artificial Intelligence Underwriting Company) announced a $40 million Series A led by Ribbit Capital, with First Harmonic participating. Total raised: $55 million.

The funding is the boring part. What AIUC sells is a standard called AIUC-1, an independent audit against it, a certification mark, and insurance tied to the result. Think SOC 2, the compliance stamp cloud companies show to prove they handle your data responsibly, but for how an AI agent behaves when someone tries to break it.

AIUC says the standard runs agents through roughly 5,000 risk-and-attack combinations covering jailbreaks, hallucinations, and data leaks, then recertifies every quarter. Cursor, ElevenLabs, Harvey, KPMG, Lovable, UiPath, and Fin already carry the mark. Cursor’s own write-up says its audit, performed by Schellman, covered secrets protection, secure code generation, MCP security, and agent identity and permissions.

Co-founder Rune Kvist, who was Anthropic’s first product hire, described the buyer’s problem in one line: most enterprises have a list of agents that passed the pilot and stalled at the security review.

I have lived that sentence from the government side. I spent years watching capable tools die in the accreditation queue, not because they didn’t work but because nobody could prove what they would refuse to do. AI agents are walking into that exact moment. Capability stopped being the bottleneck. Evidence is the bottleneck now.

What you can actually do with this

This is the part I care about most, because none of it needs a Salesforce budget.

Build a tiny Koa of your own. NVIDIA publishes Nemotron 3 Super as an open-weight download, and the post-training toolkit Salesforce used (NeMo RL and NeMo Gym) is public. You don’t need 27 years of data. Write out the ten workflows your team repeats every week, generate synthetic examples of each with a frontier model, and fine-tune a small open model on a rented GPU. Realistic expectation: a weekend for a rough version, and a model that’s worse than Claude at everything except your ten workflows, and that’s exactly what you want it to be.

Put a router in front of your AI. Even if you never train anything, you can stop sending every question to the most expensive model. A routing service or a simple rules layer in your own code can send routine tasks to a cheap model and only escalate the hard ones. Salesforce’s entire strategy is this idea with a bigger budget.

Ask for the report before you buy an agent. If a vendor says their agent is “enterprise ready,” ask whether it holds an AIUC-1 certification and whether you can see the audit scope. If they don’t, ask what independent testing covered secret exposure and MCP security. The answer tells you more than the demo.

Run your own mini inspection. Take the AIUC-1 categories and spend an hour trying to break your own agent. Paste a fake API key and see if it repeats it. Feed it a document with hidden instructions. You will find something. Everybody does. Finding it yourself beats finding it in a breach notification.

What it costs, and what you’re licensing

Koa is Salesforce-hosted and sold through Agentforce, so there is nothing to download. Salesforce has not published Koa pricing as of September 16, 2026. The pitch is that it burns fewer tokens per task than sending the same job to a frontier model, which is where the savings would come from. Availability: select pilot customers now, general availability expected winter 2026 in US regions.

The base model, Nemotron 3 Super, is a free download under NVIDIA’s open model license. Re-verify the exact terms on NVIDIA’s model page before you build a product on it. The real cost of doing this yourself is GPU rental for post-training and hosting, plus the hours to build the synthetic data.

AIUC-1 pricing is not public. Cost is driven by the independent audit (Cursor used Schellman), the adversarial testing, and quarterly recertification. Insurance is priced on top of the audit result.

Claims vs. Receipts

Claim: Koa “matches or exceeds leading model performance on CRM actions with three times fewer errors.” (Salesforce press release, September 15, 2026.) Receipt: That number comes from Salesforce’s own CRM benchmark, published in a Salesforce research paper submitted to arXiv on September 14, 2026. As of September 16, no independent lab has replicated it. Treat it as a vendor figure until someone outside the building runs the test.

Claim: TechCrunch’s headline calls Koa “everything the AI labs should fear.” (September 15, 2026.) Receipt: Salesforce signed Claudeforce with Anthropic on August 26, 2026, and Koa won’t be generally available until winter 2026. The labs are not losing a customer. They’re becoming the expensive tier in a router.

Claim: An AIUC-1 mark means an agent is trustworthy. Receipt: AIUC recertifies quarterly because agent risk changes when tools change, and a Cloud Security Alliance research note (June 5, 2026) says organizations certified against the earlier version must close new gaps before the next audit. One more wrinkle: TechCrunch’s September 15 story lists four AIUC customers; AIUC’s own release the same day lists seven. Neither is wrong. The smaller number is the one most coverage repeated.

So What

Do this: Write down your ten most repeated workflows this week. That list is the raw material for a specialized model, a router rule, and a vendor question. All of them start there.

Skip this: Waiting for Koa. It’s Salesforce-only, pilot-only, and US-only until winter.

Wait on this: Making AIUC-1 a hard purchasing requirement. It’s the strongest signal available today, but it’s one company’s standard, and the audit ecosystem is months old.

Steal this line: “Rent the genius, own the grunt work. Then get both inspected.”

Salesforce spent 27 years earning its version. You can start with ten workflows and a weekend. It’s the same recipe in a smaller kitchen.

Receipts/Sources:

The landlord and the garage

The part Salesforce isn’t saying loudly

Meanwhile, someone showed up with a clipboard

What it costs, and what you’re licensing

Claims vs. Receipts

]]>