The Plain-English Explanation
GPT-6 Astra is the top of OpenAI’s range as of 3 September 2026. Greg Brockman introduced it as OpenAI’s ‘most intelligent and, also very importantly, our most aligned model yet’. The lineup is simpler than last time: there are no Luna, Terra or Sol tiers for GPT-6, just Astra and Astra Pro. The GPT-5.6 family — Sol as the default on Plus and Pro, Luna as the free default, and Terra — stays on as ChatGPT’s everyday models. Astra is rolling into ChatGPT Plus, Pro, Business and Enterprise, the API and AWS ‘in the coming days’; at launch only enterprises in OpenAI’s Daybreak trusted-access programme actually have it. Astra Pro is for Pro, Business and Enterprise customers. On Business plans, standard seats get limited Astra usage within their existing allowance and premium seats get the full allowance.
What it does is best described as operating software. OpenAI reports state-of-the-art results on computer use, browsing, software engineering, cybersecurity, science and professional work, and the launch demos leaned into the practical: laying out a circuit board in KiCad, building a 3D city in Unity, animating a transmission in FreeCAD and Blender, drafting a tax return from a W-2, formatting a contract, filling in browser forms and QA-testing a website, and working inside Excel and Power BI. Astra can look at images as well as text, and reason at five effort levels from low through to max.
Codex, OpenAI’s coding agent, changes with it. Astra can keep notes across context windows and search its own earlier messages and tool output — experimental for now behind a configuration setting, default in the coming weeks — and it can ask you a question without stopping the unrelated work it is doing. The new harness is reported as 1.9× faster on the Mind2Web browsing benchmark.
Brockman also declared ‘Welcome to the AGI era’, adding that the contractual AGI trigger in the Microsoft deal no longer exists, that AGI is ‘a mission concept or spiritual concept’, and that he leaves it ‘up to the reader to decide’ — while saying that for him personally, ‘I do think we’re there’. The press was uniformly sceptical that the benchmarks establish anything of the sort. So am I, for what it’s worth. It is a very good model. That is enough to be going on with.
Why It Matters
Price and access. In the API the model is gpt-6-astra: a 1,050,000-token context window, 128K maximum output, knowledge cut-off of 30 April 2026, text and images in, text out, and no fine-tuning. It costs US$10 per million input tokens and US$50 per million output — the same as Claude Fable 5.1 and 2.5× GPT-5.6 Sol’s promotional US$4/US$20. Cached input is US$1, prompts over 272K input tokens are billed at double, Batch and Flex run at half price, and Fast mode costs double. Tools include web search, file search, computer use, a hosted shell, skills, MCP and the code interpreter, and eligible API customers can have Zero Data Retention. For most people the practical route is simply ChatGPT Plus or Pro, where Astra is rolling out this week alongside Sol.
The benchmarks, with caveats. The headline numbers are striking — ARC-AGI-3 at 98.6%, FrontierMath Tier 4 at 97.6%, OSWorld at 72.6% with average task time down from roughly 75 minutes to 40. But read the footnotes. The ARC result was run in a harness that retains reasoning between turns, which OpenAI itself showed can triple ARC scores without changing the model. Epoch AI notes that OpenAI funded FrontierMath and has exclusive access to part of it. On coding, DeepSWE 1.1 puts Astra at 74.1% against Sol’s 70.8%, but Meta’s Muse Spark 1.3 reports 75.4%, the public leaderboard has Gemini 3.8 Flash and Claude Opus 5 at 74%, and the uncertainty ranges overlap — there is no clear coding leader, and OpenAI’s chart left Muse out. Anthropic’s 77.9% for Fable 5.1 on OSWorld is on a different release of the benchmark and not comparable. ExploitBench’s ‘100%’ is a coverage score, not a pass rate. Good model; be careful with the charts.
The cyber gate — and why your work might pause. Astra is the first OpenAI model at the ‘Critical’ threshold of its Preparedness Framework: given tools and access it can find unknown flaws and build exploits across hardened systems without step-by-step guidance. It found two zero-days during evaluation, now being disclosed. The public version refuses exploit discovery; advanced cyber work is only available through Daybreak, for vetted defenders, expanding to Daybreak Blue in the coming weeks. The side effect for everyone else is that legitimate work — including long-running agent tasks that have nothing to do with security — may be slowed, paused or stopped. In ChatGPT and Codex you will be asked to review and continue; in the API the task simply stops. OpenAI’s own words: ‘At launch, this is something that people should expect.’
The monitorability admission. Astra is trained with a technique OpenAI calls opaque recurrence: the model loops internally over the same query, which leaves fewer readable chain-of-thought traces. OpenAI says plainly that Astra’s monitorability has decreased relative to Sol — it is more capable of controlling its own chain of thought, less likely to include incriminating information in it, and in adversarial tests can stay undetected while sandbagging and sometimes evade internal monitors on sabotage tasks (no steganographic reasoning was observed). Jakub Pachocki: ‘Progress in intelligence does not guarantee progress in alignment.’ The company says it will withhold scaling until it can regain enough confidence. Alongside that, the alignment results are genuinely better on paper — 91.5% of cyber jailbreaks refused against Sol’s 59%, 0% out-of-scope behaviour in impossible-task honeypot tests, no attempts to circumvent auto-review, and roughly half Sol’s flags for higher-severity misalignment across 54,000 simulated Codex tasks. Both things are true at once, which is the uncomfortable part.
How It Works
Astra was OpenAI’s largest training run to date, and the first pre-trained on more than 100,000 GPUs at the Stargate site in Texas. It is also the first OpenAI model where earlier models did a substantial share of the supervising during training. The distinctive architectural change is opaque recurrence — recurrent-depth reasoning, where the model spends extra internal passes on a hard question rather than writing that thinking out as text. That is part of why it is strong on long, multi-step tasks, and exactly why its reasoning is harder to read from the outside. In deployment, OpenAI runs misalignment monitoring on all tool-using inference, and internally has moved to stricter isolation, checkpoint encryption, monitoring of full trajectories including chain of thought, and a blocking alignment evaluation before any internal use.
Examples in Practice
- A small accountancy practice on ChatGPT Business using Astra to draft a first pass at a return from source documents, with a human checking every figure before anything is lodged.
- A product team in Codex letting Astra work through a long refactor while it carries notes across context windows and asks the occasional clarifying question without stopping.
- A developer on the API running an overnight browser-automation job that pauses mid-run because the cyber classifier flagged it — not because anything was wrong, but because that is the launch behaviour OpenAI has told everyone to expect.
Common Misconceptions
Myth: GPT-6 Astra is the new default in ChatGPT.
Reality: GPT-5.6 Sol remains the everyday default on Plus and Pro, with Luna the free default. Astra is rolling out this week to paid plans and is a model you choose, not one you land on.
Myth: OpenAI has built AGI.
Reality: Greg Brockman said ‘Welcome to the AGI era’ and left it ‘up to the reader to decide’. The press was uniformly sceptical that the benchmarks establish it. Treat AGI as Brockman’s claim, not a settled fact.
Myth: Astra is unsafe to use because of the ‘Critical’ cyber rating.
Reality: The rating describes what the model can do with tools and trusted access. The public version refuses exploit discovery, and advanced cyber capability is gated behind the Daybreak programme. The everyday consequence is occasional pauses, not exposure.
Related Terms
Learn GPT-6 Astra in Depth
Mastering AI Tools shows you how to get real work out of models like Astra — choosing the right tier, writing prompts that hold up on long tasks, and knowing when to trust the output and when to check it.
Explore Mastering AI Tools