OpenAI shipped its new flagship on July 9, and launch week delivered a plot twist. The British government’s AI security lab broke through the model’s cyber guardrails, and the evidence is sitting in OpenAI’s own launch paperwork.
Consider this the condensed version of multiple launch posts and one very long PDF. In the next five minutes you’ll know which of the three new models to use for what, two practical ways to wire the new Codex features into real work, and what that jailbreak finding actually changes for you.
The verdict up front: GPT-5.6 is a real upgrade and worth using today, but the win is the routing, not the flagship. Default to Terra, save Sol for work that earns it, and go straight to ultra mode in Codex, because that’s where the useful news lives. Just don’t wire any frontier model into security-sensitive work on the vendor’s word alone. This week showed exactly why.
What actually shipped
GPT-5.6 went generally available on July 9, 2026, across ChatGPT, Codex, ChatGPT Work, and the OpenAI API. The family has three tiers with a new naming system, where the number is the generation and Sol, Terra, and Luna are durable capability tiers.
Pricing per 1M tokens, straight from OpenAI: Sol runs $5 input and $30 output, Terra runs $2.50 and $15, and Luna runs $1 and $6.
Two new dials matter. A “max” reasoning effort gives Sol the most time to think, and “ultra” mode goes beyond a single agent by using subagents to accelerate complex work. In Codex, ultra is available on Plus plans and higher.
The release came with a product wave around it. ChatGPT Work, an agent for longer tasks built on Codex, can research, work across connected apps and files, and create finished documents, spreadsheets, presentations, and reports. It’s on desktop for all plans, with web and mobile rolling out to paid plans. There’s also Sites, now in public beta, for building interactive websites and lightweight apps: Business and Enterprise customers can publish publicly via URL, Pro and Edu users have beta access with Plus following, and public publishing isn’t available in the EEA, Switzerland, or the UK at launch.
The routing decision (this is the money part)
Most people will open the model picker, tap the biggest name, and run everything on Sol. That’s paying flagship prices for email drafts.
The smarter play is boring, and it works. Terra becomes your everyday default. Luna takes the high-volume simple stuff like classification and quick drafts. Sol comes out for work that actually earns it, like the debugging session that eats your whole afternoon or a long research run. You keep the quality where it matters and cut the spend where it doesn’t.
Two Codex use cases you can wire up today
The parallel refactor (ultra mode). You know the repo I mean. The legacy Python automation scripts held together with hope and comments from three jobs ago, the ones that keep surviving every cleanup sprint.
The wiring: update the ChatGPT desktop app so Codex is available inside it, set GPT-5.6 Sol as your Codex model, and toggle on ultra. Then hand it the whole job in one brief: map the dependencies, write tests for current behavior, upgrade the packages, refactor, review. OpenAI says ultra coordinates multiple agents across parallel workstreams to finish complex tasks faster, so the job you’d normally block a weekend for turns into something you supervise over lunch.
One honest caveat. Ultra and multi-agent are quality-first tools that use more compute and tokens, so don’t point this cannon at a typo fix.
Chaos in, deliverable out (ChatGPT Work). This one is for the messy middle of knowledge work, where the context lives in six places and the deliverable is due Friday.
ChatGPT Work gathers context from connected tools like Slack, Notion, Microsoft 365, and Google Drive and turns it into finished, shareable artifacts. Connect your apps and give it an outcome instead of a task. Skip “summarize this channel” and go straight to “turn this project channel, these three planning docs, and last month’s budget sheet into a client-ready status deck plus a risk tracker spreadsheet.” Then make it recurring: Scheduled Tasks can run on a schedule or a trigger, so it can independently turn new Teams and Slack messages into updated docs and flag the important changes to your team.
The jailbreak story, told straight
The UK AI Security Institute, the British government lab that safety-tests frontier models, red-teamed Sol’s cyber safeguards before release. Its team reported finding universal jailbreaks in every round of testing, meaning techniques that broadly bypass the guardrails rather than slipping past on one narrow prompt. Once through, the model could do long-form agentic work on vulnerability discovery and exploit development, and the agency said the jailbreaks were often developed within hours.
Three details keep this from being a pure horror story. First, the finding was published in OpenAI’s own system card on launch day, which is transparency worth rewarding. Second, AISI had privileged access to the model’s inner workings, including the safety monitor’s chain of thought, that real attackers wouldn’t get, though the team’s lead says the jailbreaks are likely still findable without that access, just slower. Third, OpenAI says it has worked to reproduce and mitigate the specific jailbreaks, while AISI expects further red-teaming to surface similar ones.
Here’s why this matters beyond the headline. Last month, a jailbreak that Amazon researchers reported in Anthropic’s Fable 5 prompted the US government to impose first-ever export controls that took the model offline, though Anthropic described that flaw as narrow. The controls were lifted on July 1 after two weeks of negotiation. Fortune reports AISI characterized the GPT-5.6 jailbreaks as potentially more severe, since they unlocked autonomous exploitation rather than just finding flaws. So independent safety evaluations are no longer a footnote. They’re now part of whether a model even stays online.
If you’re wiring GPT-5.6 into security-sensitive or heavily connected workflows, this week is your reminder to treat vendor safety claims as one layer of protection and verify the rest.
The rest of the fine print
The launch benchmarks are still vendor-reported. OpenAI claims 92.2% on BrowseComp, and CEO Sam Altman told CNBC the model is 54% more token efficient on agentic coding, with no independent verification of those numbers yet. In a fitting bit of launch-week timing, OpenAI itself published research arguing that one of the field’s most-cited coding benchmarks is broken, which is a useful reminder from the source that eval numbers deserve skepticism.
Early reviewers are split in an interesting way, with some testers cited by Axios seeing Anthropic’s Fable 5 as having greater raw intelligence while calling GPT-5.6 more reliable for everyday tasks.
And the pricing has layers. The headline API prices are short-context standard rates, long-context pricing runs higher, and cache writes bill at 1.25x the uncached input rate while cache reads keep the 90% discount. Run the math on your actual context lengths before any high-volume commitment.
So What
Do This: Set Terra as your everyday default and point ultra in Codex at one genuinely complex task this week to see what parallel subagents feel like.
Skip This: Running everything on Sol Pro and wiring any frontier model into security-critical workflows on vendor safety claims alone.
Wait On This: Independent benchmark evals, OpenAI’s follow-up on the jailbreak mitigations, and your long-context cost math before any big API commitment.
Steal this line for your next team meeting: route by task, not by flagship, and treat every vendor claim, benchmarks and safety alike, as unverified until someone independent checks it.