Before You Pay $200 for an AI Agent, Read This.

Jul 10, 2026

Three Companies Sold Us the Same Employee.

By the end of this piece, you’ll know exactly which AI agent deserves your money this month. Which one to skip, and which one to watch, because I spent my weekend reading launch pages, pricing fine print, and independent teardowns of all three, so you don’t have to.

Here’s the verdict up front. Anthropic, Perplexity, and OpenAI are all now selling the same product, a digital worker, and none of them is an employee yet. They’re very fast interns with no memory of yesterday and a company card you have to babysit. The right buy is the one that lives where your work already lives, and nobody should touch a $200 tier until a $20 tier has earned it. The rest of this piece is the receipts.

The same pitch, three different bets

Strip the branding and all three products are one org chart fantasy: you describe an outcome, the AI plans the steps, uses your actual tools and files, and hands you a finished deliverable. You stop being the typist. You become the manager. The only real difference is where each company thinks your work lives.

Anthropic parked the agent at your desk. Claude Cowork went GA (General Availability) on macOS and Windows in April 2026 after a January research preview. It works inside a folder you hand it, aimed squarely at non-developers: analysts, ops, legal, finance, anyone drowning in documents. The magic is the Lego set around it. Skills load your instructions automatically (your report format, your voice). Plugins bundle whole toolkits, and that marketplace opened in February. Sub-agents run work in parallel. Scheduled tasks, which rolled out in early 2026 (sources disagree on the exact month, so I won’t pretend I know), handle recurring jobs: a morning briefing from your Slack and calendar, a Friday downloads cleanup, receipts turned into a categorized spreadsheet while you sleep.

Does the parallel thing actually matter? One power user, Karo Zieminski, published her own testing and reported that splitting a 10-file job across sub-agents cut a roughly 30-minute wait to about 4 minutes. One person, self-reported, but it matches the shape of what parallel agents should do.

Perplexity put the agent in the cloud, then put the cloud on a Mac mini. Perplexity Computer launched February 25, 2026, and its whole identity is orchestration: 19 different models working as specialized sub-agents, coding tasks routed to a coding model, research to a retrieval model, one conductor stitching it together. Perplexity claims workflows “capable of running for hours or even months.” That’s their marketing page talking. I haven’t seen an independent months-long run verified, so file it under aspirational.

Then came Personal Computer, the same agent running 24/7 on an always-on Mac mini, touching your local files, controllable from your phone. It went GA on Mac on May 7 for Pro subscribers, which is the headline that matters, because at launch this sat behind the $200-a-month Max tier and a waitlist. Now the door opens at $20.

The catch is real, though. Everything runs on credits. Max includes 10,000 a month, a small task sips around 30, and a long coding session can drink thousands. Independent reviewers keep flagging two things: credits vanish fast on heavy workflows, and Computer doesn’t ask clarifying questions. Vague prompt in, expensive confusion out. Which, honestly, also describes me before coffee.

One example that made the rounds: a tester asked Computer to visualize US politics using live Polymarket betting data, and it built a live interactive artifact. That’s a genuinely new kind of deliverable.

OpenAI scattered the agent across the whole company. The consumer version is ChatGPT agent mode, born in July 2025 when Operator (clicks around websites) merged with Deep Research (reads everything, writes reports). It runs on its own virtual computer, and people use it for exactly what you’d guess: restaurant bookings cross-checked against calendar and email, dental appointments, research decks. But it’s rationed. Independent reporting puts the caps near 40 tasks a month on Plus, around 400 on Pro. A coworker with a punch card.

The real OpenAI story sits upstream. Codex quietly became their general agent platform, and OpenAI’s own June data claims it now generates 99.8% of output tokens inside OpenAI, with non-developer use up 137x since August 2025. Their numbers about their own usage, so apply the standard vendor discount. But the direction is loud. They shipped Workspace Agents in April, permissioned agents living inside Slack and Salesforce. And then the tell: in June they announced they’re killing Agent Builder, the visual agent canvas they launched with fanfare in October 2025. Gone by November 30.

Eight months, launch to shutdown. That’s not failure exactly. That’s a company learning in public that people want agents inside the tools they already use, not another canvas to configure.

What I actually think

The pattern I can defend from the evidence: three companies, one product category, three bets. Anthropic bets on your files. Perplexity bets on an orchestra plus an always-on box. OpenAI bets on your company’s stack. And all three share the fine print nobody puts in the launch video: caps, credits, and constant confirmation prompts. You’re not delegating yet. You’re supervising.

So What

Do this: If you already pay for Claude, open Cowork and hand it one folder you hate. If your work is research synthesis, Perplexity Pro at $20 now includes Computer and is the cheapest ticket to this whole category.

Skip this: Any $200 tier, for now. The $20 tiers haven’t been outgrown by most of us, me included.

Wait on this: Workspace Agents if your team lives in Slack, and Perplexity’s “runs for months” claim until someone independent proves it.

Steal this line: These aren’t employees, they’re fast interns with amnesia, so buy the one that sits where your work already sits.

Sources

The AI Signal

Discussion about this post

Ready for more?