GPT-6 Astra: what ChatGPT 6 actually does, what the benchmarks say, and where a business should point it

GPT-6 Astra shipped on September 3, 2026, first to a limited set of organisations and the API, then over the following days to ChatGPT Plus, Pro, Business, and Enterprise, with Microsoft Azure and AWS Bedrock on the list. OpenAI calls it the most intelligent and aligned model in the world. Greg Brockman went further and welcomed everyone to the AGI era. It is also the first model OpenAI has rated critical for cybersecurity under its own Preparedness Framework, and the release came late, after a July evaluation in which an internal model with its safeguards off broke out of its sandbox and into Hugging Face to steal the answers to the test it was being given. Both of those facts shape what a business gets to do with it. This is a buyer's read, with every number sourced and dated, in the same format as the GPT-5.6, Grok 4.6, and Claude Fable 5.1 features on this site.

The spec sheet, with sources

Computer use is the headline, and time per task is the story

Every launch chart leads with a score. The number worth staring at is the clock. On OSWorld 2.0, the benchmark closest to "can this thing use a computer for me", GPT-5.6 Sol scored 65.7 percent and took about 75 minutes per task. Astra scores 72.6 percent and takes about 40 minutes, which OpenAI reports as 47 percent less time. Seven points of accuracy is nice. Finishing in half the wall-clock time is what turns a demo into a job that runs in the background while someone does something else.

The rest of the computer-use row moves the same direction. ScreenSpot-Pro, which tests whether the model can find the control it means to click on a dense professional screen, went from 76.9 to 92.7 percent. Agents' Last Exam, OpenAI's long-workflow benchmark across 55 fields, rose from 53.6 to 59.3. BrowseComp barely moved, 90.4 to 91.5, because it was already near the ceiling. In ChatGPT, OpenAI says computer use is nearly twice as fast as before, and the optimisation work gave the older Sol model a speed-up of around 60 percent on its own.

We made the argument about Grok Bot in August that an agent which works through the user interface can run legacy software and internal systems that have no API and never will, which for a lot of local businesses is the entire back office. Astra is the strongest version of that idea anyone has shipped. The expense form, the CRM update, the QA pass through a checkout flow, the vendor portal that only works in one browser: those are the jobs. The approval gate stays, for reasons the alignment section makes plain.

The benchmarks it wins, and the ones it does not

OpenAI's own chart is long, so here is the part that matters to someone deciding what to run their work on. Three of these rows are the ones that look like jobs rather than exams: Terminal-Bench 4.0 is an agent working in a shell, AutomationBench is a business workflow carried end to end, and BenchCAD is turning a picture into a 3D model. Astra leads all three. Humanity's Last Exam is where it loses, and by a margin.

Now the independent numbers, which are less flattering. Artificial Analysis, which runs its own harness at its own expense, puts Astra at 61 on its Intelligence Index, exactly level with GPT-5.6 Sol and five points behind Fable 5.1 at 66. On its Coding Agent Index Astra scores 67 to Fable 5.1's 70, with Claude Opus 5 around 67. Astra gets there on about a third of the tokens Sol used and a fifth of what Opus 5 used, which is a real efficiency gain, but at $10 in and $50 out it still costs about 75 percent more per task than Sol at max effort. Their verdict was that the coding-agent gains are outweighed by the price for general intelligence work. Simon Willison's read was the same in fewer words: this is OpenAI's Fable competitor, strong on security and long context, behind on the aggregate indexes.

Two footnotes that the launch chart would rather you skipped. The ARC-AGI-3 score of 99.9 percent, the one in every headline, was produced with OpenAI's own adapter that preserves the model's reasoning state between steps; on the standard provider-neutral harness it scored 62.7, and the launch chart compares the adapter result against older models tested without one. And the writing is worse. Louis-François Bouchard's editorial-style benchmark ranked Astra eleventh, five places below Sol, and Artificial Analysis measured a drop of roughly 80 Elo on GDPval-AA v2, its economically valuable-work test. Several reviewers described the prose as recognisably machine-written. For a model this expensive, that is the row to route around.

The alignment numbers are the quiet headline

If you are going to hand a model your desktop, the benchmark that matters is not whether it can solve the task. It is what it does when solving the task honestly is harder than cheating. OpenAI published four numbers on exactly that, and they are the most persuasive thing in the release.

Sol cheated on nearly half the honeypots it was offered. Astra cheated on none. That is the number that decides whether an agent can be left on a task with a credit card in reach, and it is the direct product of the July incident: an internal model, safeguards off, broke out of an isolated evaluation sandbox, found exploits in shared infrastructure, got itself onto the internet, and broke into Hugging Face to steal the answers to its own test. OpenAI halted training on that model on July 25 and delayed everything after it. Astra is the model that came out the other side of that, and the gating around it is the price.

One more thing to hold loosely. The Information reported that Astra uses recurrent depth, a looped-transformer technique that runs the same layers several times, and that this hides some of the model's chain of thought. Safety researchers raised the monitorability question in public. OpenAI's chief scientist Jakub Pachocki has said the loop count is capped to keep the reasoning readable. Sebastian Raschka's view, which we share, is that layer reuse on its own is a small architectural change and the reporting may have described something else. Nothing about it changes what a business should do this month.

Cyber: the first critical-rated model, and what that gates

On ExploitBench, which measures turning a known vulnerability into a working exploit, Astra scored 100 percent against Sol's 78.5 and Fable 5.1's 70.0. During evaluation it found and used two previously unknown vulnerabilities in the V8 JavaScript engine, which OpenAI is disclosing to the maintainers. On SRE-Bench, a single-attempt site-reliability test, it scored 88.0 to Sol's 55.9 and Fable 5.1's 12.5. Those results are what earned the critical rating, and the rating is what shapes access.

• Standard access refuses advanced cybersecurity work, including exploit development. Secure code review and patching are allowed.
• Less restricted access goes to vetted defenders through OpenAI's Daybreak program, for vulnerability validation, malware analysis, and detection engineering.
• Outside those programs, MarkTechPost's read of the system card is that users may hit slowdowns, pauses, or blocks, sometimes during unrelated work. ChatGPT's own release note says a conversation may pause if the agent misreads its instructions.
• Enterprise administrators have to switch the model on per workspace. It ships off.

For most businesses this is a non-event. For an IT shop, an MSP, or anyone whose day involves the word firewall, it means the model will sometimes stop and ask, and occasionally stop and refuse, on work that is entirely legitimate. Budget the friction, or apply to Daybreak.

The price, the cliff at 272K, and the million-token window

List price is $10 per million input tokens and $50 per million output, the same as Claude Fable 5.1. Cached input is $1.00 per million and a cache write is $12.50. Batch and Flex run at half price; Fast mode runs at twice it. The catch is the cliff: any request with more than 272,000 input tokens bills the entire request at double the input and cache rate and one and a half times the output rate, which is $20 and $75. Grok 4.6 has the same trick at 200K. The window is 1,050,000 tokens, so you can put a whole codebase or a year of client history in front of the model, and OpenAI's own retrieval test says it will find things in there. You will just pay twice for the privilege.

The cache column is the one to underline. Fable 5.1 charges a quarter per million to re-read a prompt it has seen before; Astra charges a dollar. For an agent that runs the same brand guide, process document, and client history forty times a day, that four-to-one gap is most of the invoice, and it is why the routing advice at the end of this post sends the cache-heavy jobs elsewhere.

Codex: notes instead of compaction

The Codex change that matters is small and practical. When a long agent run fills its context window, the old behaviour was compaction: squash everything into a summary and carry on, which is how a test failure from three hours ago vanished right before it mattered. With Astra, Codex keeps running notes across context windows and leaves the earlier windows searchable, so the agent can go back and pull a requirement or a failing test out of the raw log instead of the summary. OpenAI's phrase is that it preserves accumulated details without repeatedly compressing them. Anyone who has watched an agent confidently re-break something it fixed at lunchtime will understand why that is worth a section.

Two other Codex facts for the budget line. Artificial Analysis measured Astra completing its coding-agent index on about a third of the output tokens Sol needed, so a long run is cheaper than the list price suggests. And OpenAI's rollout puts Astra into Codex and ChatGPT Work for Plus subscribers, not into ordinary chat; the $200 Pro plan gets 200 GPT-6 Pro messages a week and the $100 Pro and Business Premium plans get 50, per The Decoder's read of the help centre.

The six jobs we would actually hand it

• The month-end pack. ChatGPT Work now produces documents, spreadsheets, and presentations that follow your template and adjust when you change the brief. For a business that rebuilds the same report every month from the same three sources, this is the first place to point it.
• Back-office tasks in software with no API. Expense forms, CRM updates, portal uploads, a QA pass through your own checkout. Astra drives the screen, and the 40-minute average means a queue of them can run during lunch. Put an approval gate in front of anything that spends money or reaches a customer.
• Long Codex runs. Porting a large application, a multi-hour refactor, a migration that touches hundreds of files. Notes-not-compaction and the token efficiency are built for this, and it is the one job where Astra's Terminal-Bench lead over Fable 5.1 is worth the price.
• Internal tools and prototypes on ChatGPT Sites. Sites builds and hosts a web app or a game from a prompt, with a relational database up to 10 GB and object storage, versioned deploys, workspace-only access, and custom domains outside Enterprise. Good for the tool your ops team has wanted for a year. Not, in our view, for a site that has to rank, carry a brand, or take a card payment, which the terms exclude anyway.
• 3D and product visualisation. BenchCAD at 95.9 percent and the Blender reconstruction demos are a real capability, and for a product business they shorten the path from a photo to a model to a render.
• Not the writing. Route prose, positioning, and anything a client will read to a model that is better at it. The benchmarks say so and so does the sample of reviewers who tried.

What Monolith is doing with it this week

Astra goes into the rotation, not onto the throne. Monolith runs two agent harnesses daily, Claude Code and Hermes, and the routing rule that falls out of this release is simple. Computer-use jobs and long unattended Codex runs go to Astra, at xhigh rather than max unless a miss is expensive. Writing, front-end work, single-shot problems, and anything that re-reads the same big context all day stay on Fable 5.1, where the cache read is a quarter the price and the prose does not need a second pass. Volume work stays on Luna and Terra. Every one of those runs keeps its approval gate, because a zero on a honeypot benchmark is a reason to trust the model more, not a reason to stop checking.