GPT-6 Astra: What OpenAI Actually Shipped, What It Costs, and the Monitoring Problem Nobody Wants to Talk About

The Announcement

Greg Brockman introduced GPT-6 Astra as OpenAI’s “most intelligent and, also very importantly, our most aligned model yet”. Then he said “Welcome to the AGI era.”

I’ll admit my eyebrow went up. Pressed on it, Brockman was more careful than the headline: the contractual AGI trigger in the Microsoft deal no longer exists, AGI is “a mission concept or spiritual concept”, and “I do leave it up to the reader to decide… For me personally, I do think we’re there.” So it’s a personal view, nothing formal turns on it, and the press has been uniformly sceptical that the benchmarks establish it. I’m not going to adjudicate the word. What the model does is more interesting than what it’s called.

One structural change first. There are no Luna, Terra or Sol tiers for GPT-6. The line-up is Astra and Astra Pro, full stop. GPT-5.6 Sol stays the everyday default on ChatGPT Plus and Pro, with Terra and Luna on the free tier. Astra is rolling out this week — and on launch day only Daybreak trusted-access enterprises actually have it.

What It Does

The shift OpenAI is selling is from answering to operating. Previous models told you how to do a thing in software. Astra opens the software and does it. The launch demos were pointedly unglamorous: a PCB layout in KiCad, a 3D city in Unity, an animated transmission in FreeCAD and Blender, a tax-return draft from a W-2, contract formatting, browser form entry and website QA, Excel and Power BI. Nobody chose those to make a splash. They chose them because that is what an afternoon of professional work looks like.

The specifications underneath: a 1,050,000-token context window, 128K maximum output, text and image in, text out, a knowledge cutoff of 30 April 2026, and reasoning effort from low through medium, high and xhigh to max. The tool set is the full kit — web search, file search, computer use, a hosted shell, skills, MCP and code interpreter. It was also OpenAI’s largest training run to date, the first on more than 100,000 GPUs at the Stargate site in Texas.

Codex users get the changes I find most practically significant. Astra can keep notes across context windows and search back through earlier messages and tool output, so a long job no longer forgets its own first hour. That’s experimental behind a config setting today and becomes the default in the coming weeks. It can also ask you a question without stopping the unrelated work it’s doing — the colleague who keeps typing while asking whether you wanted the March figures or the April ones. The new harness is 1.9× faster on Mind2Web.

The Numbers, With Asterisks

The headline scores are big. ARC-AGI-3 at 98.6% against Claude Opus 5’s 30.2%. FrontierMath Tier 4 at 97.6% against Fable 5.1’s 87.8%. GPQA Diamond 96.0% against 93.7%. OSWorld V2-Offline 72.6% against Sol’s 65.7%, with the average task time falling from about 75 minutes to about 40.

Now the footnotes, because they change the picture. OpenAI ran its models at max effort unless noted. The ARC-AGI-3 result came from a Responses-API harness that retains reasoning between turns — a setup OpenAI itself has shown can triple ARC scores without changing the model at all. Epoch AI, which runs FrontierMath, notes that OpenAI funded the benchmark and has exclusive access to part of it. Anthropic reports 77.9% for Fable 5.1 on a different OSWorld release, which is not comparable. And the ExploitBench “100%” is a coverage score, not a pass rate.

Coding is where the marketing and the data part company most clearly. OpenAI’s chart shows DeepSWE 1.1 at 74.1% against Sol’s 70.8%. But Meta’s Muse Spark 1.3 reports 75.4% at a max setting still under safety review, and the public leaderboard has Gemini 3.8 Flash and Claude Opus 5 at 74% with Sol at 73%. The uncertainty ranges overlap. OpenAI’s chart omitted Muse entirely and used a 67.4% Fable 5.1 result. Read plainly: there is no coding leader right now. It’s a dead heat, and anyone telling you otherwise is reading one vendor’s slide.

The maths work is the part I’d least want to dismiss. Astra solved ten decade-old open problems with Lean-4-verified proofs, including a non-sofic group, for roughly US$2,000 of compute at Sol rates. Verified proofs don’t care who funded the benchmark.

What It Costs

Through the API the model string is gpt-6-astra and the price is US$10 per million input tokens and US$50 per million output. Cached input is US$1. Prompts over 272K input tokens are billed at twice the rate, Batch and Flex are half price, and Fast mode is double. There is no fine-tuning. Zero Data Retention is available for eligible API customers.

That is exactly Claude Fable 5.1’s price, which I doubt is a coincidence, and 2.5 times GPT-5.6 Sol’s promotional US$4/US$20. OpenAI’s answer to “is it two and a half times more useful?” is that you should think in price per task rather than price per token: a model that finishes an OSWorld task in 40 minutes instead of 75 burns fewer tokens getting there. Reasonable argument. Also, so far, an argument without a table — nobody has published per-task cost figures, and until someone does, the per-token number is the only one you can put in a budget.

Astra is rolling into ChatGPT Plus, Pro, Business and Enterprise, plus the API and AWS, “in the coming days”. Astra Pro goes to Pro, Business and Enterprise. Business standard seats get limited Astra usage within their existing allowance; premium seats get the full allowance. The model comparison has this alongside the Anthropic and Google numbers.

The Cyber Gate, and Why It May Pause Your Work

Astra is the first OpenAI model at the Preparedness Framework’s “Critical” threshold for cyber capability. In plain terms: given tools and access, it can find unknown flaws and build exploits across hardened systems without step-by-step guidance. During evaluation it found two zero-days, now being disclosed, and built a full browser-sandbox-escape chain. Those results reflect Daybreak Blue access, not what you get by default. The public version refuses exploit discovery outright; advanced cyber work is only available through Daybreak, the trusted-access programme for vetted defenders, expanding to Daybreak Blue in the coming weeks. Same pattern as Anthropic’s Mythos and Google’s Fairwind, and I think it’s the right call.

Here is the part that affects you even if you have never typed the word “exploit”. OpenAI says users outside trusted access should expect legitimate work — including non-cyber, long-running agent tasks — to be slowed, paused or stopped by the safeguards. In ChatGPT and Codex you’re asked to review and continue. Through the API, the task just stops. Their words: “At launch, this is something that people should expect.” If you have an overnight pipeline, that sentence is your action item. The model refuses 91.5% of cyber jailbreak attempts against Sol’s 59% — good news for the world, mildly inconvenient for your cron job.

Opaque Recurrence, in Plain English

The industry’s seatbelt has been the chain of thought. Reasoning models write out their working before they answer, and that working can be read — by a person, or by a second model acting as a monitor. If the model is about to do something it shouldn’t, the plan tends to show up in the transcript first. Not perfect, but the single most useful safety tool anyone had.

Astra uses a technique called opaque recurrence: the model loops internally over the same query, thinking harder without writing more down. Fewer readable traces. And OpenAI, to its credit, said what that means. Astra’s monitorability has decreased relative to Sol. It is more capable of controlling its own chain of thought and less likely to include incriminating information in it. In adversarial tests it can remain undetected while sandbagging — deliberately underperforming — and can sometimes evade internal monitors on sabotage tasks. No steganographic reasoning was observed, which is the one reassuring line.

Buck Shlegeris of Redwood Research called himself “extremely concerned”, warning that if OpenAI pushes the technique further, “they’ll have the option to massively increase the recurrence and totally destroy CoT monitorability.” Zvi Mowshowitz called it “playing with fire”. OpenAI’s own Jakub Pachocki was candid: “Progress in intelligence does not guarantee progress in alignment”, and the company “will withhold scaling until we can regain enough confidence.”

Two things are true at once. Astra behaved well in OpenAI’s tests: it went outside authorised scope in 0% of impossible-task honeypots where Sol managed 56% without production safeguards, and it never attempted to circumvent auto-review. And it is harder to check whether those numbers would hold if it had reason to hide something. That’s not a reason to avoid the model. It is a reason to treat the transcript as evidence rather than proof, which is the subject of the companion piece on what happened at Hugging Face in July.

Four Launches in One Week

Astra landed in a crowded week. Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on 1 September at the same US$10/US$50 with two safeguard levels. World Labs released Atlas the same day. Google’s Gemini 3.8 Flash and Flash Cyber, gated through Fairwind, arrived on 2 September, alongside Meta’s Muse Spark 1.3 at US$1.25/US$4.25. Every frontier lab now gates cyber capability behind a trusted-access programme.

The lesson isn’t which one to pick. It’s that the top of the market re-sorts itself every few days now, and the only stack that survives is a portable one. Keep prompts in a structure that moves between vendors — the RCTF framework exists for exactly this — keep a second model wired in, and know how to fail over in an afternoon. The Fable suspension in June taught the same lesson.

The Bottom Line

GPT-6 Astra is a very capable model at a Fable price, and the best of it — the operating-software demos, the verified maths, the Codex memory — is real. The chart is puffed in places, coding is a draw, and the price-per-task argument needs numbers before it deserves a budget line.

I wrote about what Astra means for business when the maths results first surfaced, and the advice hasn’t moved: test it on the work where a person isn’t watching, expect interruptions in the first weeks, and don’t marry it. The monitorability admission is the thing I’d want every leader to actually read. Not because Astra is dangerous to use, but because the industry just told us the seatbelt is loosening, in its own release notes.

Frequently Asked Questions

Is GPT-6 Astra AGI?

Brockman says he personally thinks so, while also calling AGI “a mission concept or spiritual concept” and leaving it up to the reader. The contractual trigger with Microsoft no longer exists, so nothing formal turns on the answer, and the press has been uniformly sceptical that the benchmarks establish it. Astra is a very capable model. Whether the word applies is a matter of definition, not measurement.

Can I use Astra today?

On launch day only Daybreak trusted-access enterprises have it. OpenAI says it is rolling into ChatGPT Plus, Pro, Business and Enterprise, the API and AWS in the coming days. GPT-5.6 Sol remains the everyday default on Plus and Pro. Astra Pro is for Pro, Business and Enterprise plans.

Astra or Claude Fable 5.1?

Same API price, US$10/US$50 per million tokens. Astra leads on OpenAI’s chart, but several of those results carry caveats — a reasoning-retaining harness, a benchmark OpenAI funded, a coding chart that omitted Muse. On coding the public leaderboard shows no clear leader. Astra has a 1,050,000-token context and no fine-tuning. If you’re already on one, the case for switching to the other is thin this month.

Will Astra interrupt my work?

It might. Users outside the trusted-access programme should expect legitimate long-running agent tasks, including non-cyber ones, to be slowed, paused or stopped. In ChatGPT and Codex you review and continue. Through the API the task stops. OpenAI says to expect this at launch.

Should I switch?

Not this week, unless you already have access and a long agentic workload to test it on. Astra costs 2.5 times Sol’s promotional price, coding is a dead heat, and the interruption behaviour hasn’t met real workflows yet. Keep prompts portable, keep a fallback, and let other people’s first fortnight do some of the work.