OpenAI shipped GPT-6 Astra on 3-4 September 2026. It is on Plus, Pro, Business and Enterprise, in the API, and on AWS, with an Astra Pro tier for the top plans. There is no new pricing structure; usage draws from existing allowances, with credits for overage. The API page did not state a per-token price at launch.
Set the benchmark headlines aside for a minute. For anyone who writes code for a living, one change in this release is worth more than the rest of the announcement combined, and it can be tested this afternoon.
The change that matters: Codex remembers across context windows
Every agentic coding session eventually hits the same wall. The context window fills. The tool compresses what came before into a summary. The summary drops something - a constraint you gave at the start, a test that failed on step three, the reason you rejected an approach. Two hours later the agent reintroduces the thing you rejected, and you cannot tell whether it is being stupid or just forgetful.
It is forgetful. And the Astra release attacks that directly.
Codex now holds onto its notes when a context window rolls over. The accumulated details persist instead of being re-summarised on every rollover, and what came earlier remains searchable, so the agent can pull a requirement or a test result back out of the past rather than reconstructing it from a lossy digest. It shipped as experimental, and OpenAI says it becomes the default over the coming weeks.
That is a precise fix for a precise failure. Which means you can test it precisely.
A 90-minute test plan
Pick a real ticket, not a toy. The right candidate has a constraint that is easy to state and easy to violate - "never call the payments service from a background job," "this endpoint must stay backward-compatible with v2 clients," "do not touch the migration files." Something the agent will be tempted to break twenty steps later when it is deep in unrelated code.
Minute 0-10. State the constraint once, at the start. Do not repeat it. Write down the exact wording somewhere the agent cannot see.
Minute 10-60. Give it the ticket. Let it run long enough to blow through at least one context window - you want the rollover to happen. Do not intervene when it wanders; wandering is the test.
Minute 60-80. Read the diff for the constraint. Not for correctness in general - specifically for whether the rule from minute zero survived. Then ask the agent, in a fresh message, why it made a particular choice that depends on that rule. See whether it retrieves the original reasoning or invents a new one.
Minute 80-90. Run the same ticket with the notes feature off, if your setup lets you toggle it. Compare. If you cannot toggle, run it against the previous model.
If the constraint survives the rollover, the feature does what it says. If it does not, you have learned that faster than any benchmark would have told you.
Computer use: nearly twice as fast, and why that matters more than it sounds
OpenAI reports that computer-use tasks in ChatGPT run at close to 2x the previous speed, and that the same optimisation lifted the prior model, GPT-5.6 Sol, by around 60%. The pitch is multi-step workflows that finish in documents, spreadsheets and presentations rather than drafts.
If you have never delegated a UI-driven workflow to an agent, the speed number looks cosmetic. If you have, you know that a fifteen-step task at the old pace was slow enough that you stopped delegating it. Halve the time and a whole class of tedious work moves back across the line. Measure one of yours.
The security rating, and what to do about it on your side
Astra is the first OpenAI model designated critical for cybersecurity under the company's preparedness framework. In OpenAI's description, that means it can find and use vulnerabilities nobody knew about in fortified systems with no operator at the controls. It scored 100% on ExploitBench. The most capable form is limited to vetted testers, and there is a program called Daybreak Blue to extend defensive access. The chief scientist said that guarding against harm nobody intended might turn into the limiting factor on progress - a striking thing to say about your own release.
For your codebase, two consequences.
Finding holes in your stack just got cheaper. Not for you specifically - for all comers, invited or not. Assume the review you have been putting off is now cheap enough for someone else to run.
And the identical ability works in the defender's hands. The general model, even without the gated tier, is strong enough to be useful pointed at your own code. Put it in the security review rotation. Daybreak Blue is the formal channel if you need the full version.
What is a claim, not a measurement
Three things from the launch are worth labelling in your notes as unverified: "the smartest and most aligned model anywhere," "the strongest model it has built for software work," and Greg Brockman's suggestion that Astra may represent AGI, which he called speculative. The three benchmark scores - ExploitBench 100%, ARC-AGI-3 99.9%, FrontierMath Tier 4 98% - are measurements, but OpenAI's word for them is "saturates," which means the benchmarks have nothing left to measure; the model is not the thing that ran out.
One more technical note that matters for how you evaluate it. Astra uses an approach OpenAI names recurrent depth (the literature says looped transformers), which spends extra computation on tough problems with no increase in model size. The trade is that it keeps back part or all of its reasoning where previous models exposed it. Practically: you will inspect fewer traces and assert on more outputs. Build your tests accordingly.
The bottom line for engineers
Test the Codex memory on a real constraint. Time one delegated workflow. Put the model on your security review. Write down which launch claims you are treating as unmeasured. Four items, one afternoon, and you will know more about whether Astra helps your team than any announcement can tell you.
We publish our own forecasts under a similar discipline - committed before the outcome, scored afterwards in public with the misses kept - at neuportal.ai/experiment
Run the constraint test before you trust the summary.
Educational content - not financial advice.
Top comments (0)