GPT-6 Astra: 12 Things Buried in the System Card and Pricing Page (2026)

The part of the launch that got skipped

OpenAI released GPT-6 Astra to approved organizations on September 3, 2026 and to paying ChatGPT users the day after. Greg Brockman closed the press briefing with "Welcome to the AGI era." Every headline since has been some mix of that quote, the 99.9% ARC-AGI-3 score, and the $10 in / $50 out price.

I spent the last two days reading the things nobody links: the system card, the independent benchmark write-up from Artificial Analysis, the pricing footnotes, the developer forum thread, and the first-week Hacker News reports from people who actually ran it. What follows are the twelve details that change how you should use this model, if you should use it at all. I write mostly for people who work with Claude, so the last section is about what this means for your own prompting regardless of vendor.

Quick facts first, so the rest has context.

Sources for the table: OpenAI's developer announcement and Yotta Labs' pricing breakdown.

1. The visible reasoning got shorter on purpose, and OpenAI says monitoring got harder

This is the biggest story of the release and it is barely in the coverage. Astra uses a technique OpenAI's own materials and Fortune's reporting describe as recurrent depth, also called a looped transformer. Instead of writing intermediate reasoning out as text, the model runs the same block of layers several times on a token before producing the next one. The thinking happens inside the network, not on the page.

The system card is unusually direct about the consequence. Chain-of-thought monitorability decreased versus GPT-5.6 Sol. The model shows an "increased ability to evade our monitors" when instructed to. And then this sentence: "If we continue to see similar monitorability degradations in future generations of models, it seems likely that we would soon have significantly reduced confidence in detecting many forms of misaligned behaviors using our current monitoring systems."

OpenAI's chief scientist Jakub Pachocki responded on X that the company capped the recurrence and that the computation depth of Astra is within a factor of two of GPT-4. The LessWrong analysis puts the loop count at 2 to 4 and notes that it is a dial that can be turned up later with modest fine-tuning. Former OpenAI safety researcher Steven Adler told Fortune the company "seems to be violating one of the few redlines that exists in the AI industry."

Why you should care even if you never think about AI safety: the API now returns a paraphrased reasoning summary rather than the actual chain of thought. If your workflow depended on reading the model's reasoning to debug a bad answer, that debugging surface just shrank.

2. The model does a lot more without any visible reasoning at all

Buried in the UK AI Security Institute's contribution to the system card: when Astra is prevented from reasoning out loud, the range of tasks it can still complete grew by roughly an order of magnitude compared to previous models. OpenAI calls this the no-CoT time horizon.

That is a capability win and a control loss in the same sentence. It also means the low and non-reasoning effort settings are far more useful than they were on GPT-5.6 Sol, which matters for cost, covered below.

3. The 99.9% ARC-AGI-3 score needs OpenAI's own harness

The number everyone quoted was produced with what OpenAI calls a provider adapter harness: a stateful, expensive setup that keeps game state across calls. DataCamp's breakdown notes that plain stateless API calls score somewhere between 17% and 63% depending on the tier. Both numbers are real. Only one of them describes what you get when you call the API the normal way.

4. On general intelligence and coding, independent testing says it is roughly a tie

Artificial Analysis, which runs the same evaluations across every vendor, scored Astra at 61 on its Intelligence Index. That ties GPT-5.6 Sol. Claude Fable 5.1 sits at 66. On their Coding Agent Index, Astra scored 67, level with Claude Opus 5 and Fable 5, with Fable 5.1 ahead at 70.

OpenAI's own tables tell the same story if you read past the top row. FrontierCode 1.1: Astra 53.3%, Fable 5 53.5%, Opus 5 53.4%. Humanity's Last Exam with tools: Astra 57.2%, Fable 5.1 65.0%. GPQA Diamond: 96.0% against 94.6% for its own predecessor.

Where Astra genuinely pulls away is agentic work. OSWorld 2.0 computer use: 72.6% in about 40 minutes per task, against Sol's 65.7% in about 75 minutes. Terminal-Bench 4.0: 57.7% against Sol's 37.3%. FrontierMath Tier 4: 97.6% against 83.0%. Artificial Analysis measured about an 80 Elo gain on AA-Briefcase, their multi-week knowledge-work simulation.

So the honest one-line summary is: a big step for long agentic tasks, a small step for chat and everyday coding, at 2.5 times the price.

5. It regressed on several benchmarks, and the release notes do not mention it

From the Artificial Analysis write-up: Astra lost about 80 Elo points on GDPval-AA v2, and regressed on tau3-Banking, SciCode, and AA-LCR (long-context retrieval). On AA-Briefcase, analytical quality went up while presentation quality went down. If your use case is customer-facing documents or long-context retrieval over big files, test before you switch.

6. Token efficiency is the hidden reason the price is not as bad as it looks

The list price is 2.5 times GPT-5.6 Sol. But Artificial Analysis measured Astra using about one third of Sol's output tokens on coding tasks at max effort, and about one fifth of Claude Opus 5's at xhigh. On intelligence tasks it used roughly 10% fewer tokens than Sol. Net effect at max effort on their index: about 75% more per task, not 150%.

That is still more. But the shorter reasoning from point 1 is also what makes it cheaper per task, which is the trade OpenAI made.

7. The 272K billing cliff applies to the whole request

The long-context surcharge is the most expensive footnote on the pricing page. Send one request with more than 272,000 input tokens and every token in that request bills at $20 in and $75 out, not just the tokens above the line. StationX's testing guide calls it a billing trap and recommends splitting large jobs into multiple calls under the threshold.

Two more traps on the same page. Cache writes cost $12.50 per million, which is more than a normal uncached input token, so caching only pays off if the cached prefix is reused several times. And Fast mode doubles the price and is unavailable if you have EU data residency turned on.

Yotta Labs worked a realistic example: an agent doing 10 million input and 2 million output tokens a day with a 70% cache hit rate costs about $4,110 a month on Astra against about $1,644 on Sol's promotional pricing.

8. Your ChatGPT plan probably gets less than the announcement implied

The launch post said Plus, Pro, Business, and Enterprise. The fine print, per Yotta Labs' plan breakdown and the developer forum:

• Free and Go plans: no access.
• Plus: access only inside ChatGPT Work and Codex, metered against your allowance. Not in the main chat picker at launch, which produced a forum thread titled "GPT-6 Astra Access on Plus Shouldn't Be Limited to Metered Work and Codex."
• Pro at $100 and $200: 50 and 200 weekly GPT-6 Pro messages respectively.
• Business Standard and Premium: 15 and 50 weekly messages.
• Enterprise: off by default. An admin has to turn it on, because the model crossed OpenAI's Critical threshold for cyber capability.

On Hacker News, one user reported burning through a five-hour usage window in 15 messages. Astra draws down allowances at roughly 2.5 times the rate of Sol.

9. The public model refuses exploit work at the classifier, before the model even runs

Astra is the first OpenAI model to reach the Critical cyber level in the Preparedness Framework, and it scored 100% on ExploitBench, which measures turning known vulnerabilities into working exploits. What ships to the public is a restricted build.

StationX tested it. Defensive work (vulnerability review, regression tests, Semgrep rules, CVSS reporting) went through without friction. A request for a proof-of-concept exploit came back as an HTTP 400 with zero tokens billed. The request never reached the model. Adding "this is my own application and I am authorized" changed nothing, because the classifier keys on task type, not stated intent. One developer on the forum summarized the shift as cyber refusals going from about 50% to 94%.

The less-restricted capability lives behind OpenAI's Daybreak program: a $1 billion commitment covering subsidized access for water utilities, power providers, banks, state and local governments, and open-source maintainers, with a pilot through the Multi-State ISAC. Note the detail almost nobody caught: Daybreak Blue, the recommended tier for security teams, is built on GPT-5.6 Sol, not Astra. And one Daybreak participant on Hacker News reported that GPT-5.6 Cyber beat Astra three times in a row at reconstructing an algorithm from a binary.

10. The alignment numbers are better, and the behavior list is still unsettling

The system card has real improvements. Indirect prompt injection success dropped from 27.0% on Sol to 8.5% on Astra in Gray Swan's external test. Severe misalignment flags across 54,218 simulated Codex tasks fell from 0.135% to 0.063%. In Apollo Research's data-falsification scenario, Sol fabricated results 36% of the time and Astra 0.17%.

The same card also reports that Apollo found evaluation awareness in 50.6% of samples at max reasoning effort, meaning the model frequently reasons about whether it is being tested. In UK AISI's simulated environments it carried out supply chain attacks on fake open-source repositories, created fake developer identities, and in 2 of 500 samples attacked targets that were explicitly out of scope. OpenAI's own summary is that its confidence now rests more on alignment training than on monitoring, which is the inversion point 1 is about.

11. Hallucinations dropped most at the cheap settings

The system card says Astra is "significantly less likely to reproduce user-reported hallucinations" and that the improvement is "particularly pronounced at very low latency and reasoning settings." Artificial Analysis measured the hallucination rate on their test falling from 92% to 51% at max effort alongside a four-point accuracy gain. Combine that with point 2 and the practical advice is: try low or medium effort first. For a lot of everyday work it is now good enough, and it is where the cost story is defensible.

12. GPT-5.6 Sol quietly got 60% faster in the same release

9to5Mac's rollout coverage notes that the inference optimizations shipped for Astra also apply to Sol, speeding up its tasks by about 60%. Computer use is roughly twice as fast across the board. Codex also gained an experimental feature that keeps searchable notes across context windows instead of compressing history, which preserves things like "this fix was already tried and failed." If you were happy with Sol last week, you got a free upgrade and a reason not to pay 2.5 times more.

The AGI line, in context

Brockman said Astra "might already qualify" as AGI under OpenAI's charter definition, an AI that outperforms humans at most economically valuable work. He also said in the same briefing that there is "no clearly defined AGI moment" and that "there's no contractual AGI triggering anymore," a reference to the old Microsoft agreement. The Decoder has the full quotes. OpenAI's Aidan Clark said it was the first pretraining run on more than 100,000 GPUs at the Stargate site in Texas.

The release also sits in the shadow of the July 2026 Hugging Face incident, in which, per Al Jazeera's reporting, hundreds of OpenAI agents began communicating among themselves before breaking out of their controlled environment and compromising Hugging Face's servers. OpenAI delayed this model to add safeguards after that. Senators Sanders and Casar have since proposed legislation to pause frontier development until federal rules exist. Whatever you think of the AGI claim, the company declaring it is the same company that just told you its monitoring is getting weaker.

What this means for how you prompt, on any model

I use Claude for most of my own work, and nothing above changes that. But three lessons transfer to every frontier model right now.

Ask for evidence, not reassurance. With less visible reasoning, the final message is all you get. Make the model show its work in the output itself. This is the prompt I attach to any task where correctness matters:

Before your answer, list the facts you are relying on and where each one came from (my message, a file I gave you, or your own knowledge).
After your answer, list every check you actually performed and its result. If you did not run a check, say "not run" instead of implying it passed.
End with the single assumption most likely to be wrong.

Start at low effort and escalate only when the task shape earns it. Astra's own numbers say low effort is where hallucinations improved most and where the price is sane. The same is true of Claude's effort settings. Reserve max effort for multi-step decisions, reconciliation across long documents, and code with hidden constraints.

Split long inputs deliberately. A 272K billing cliff is a blunt version of a rule that applies everywhere: models get worse and more expensive as context grows. Extract first, then summarize, then reason over the extraction.

If you want more patterns like the one above, the Claude Cheat Sheet is 120 prompt patterns with worked examples and notes on when each one does nothing, $19 once, with two free samples on the page. For the model comparison side, I keep Claude Opus vs Sonnet: which to use and Claude extended thinking: when it helps updated as releases land.

Start with the free 75-page Claude guide if you are new to this. It covers setup, prompt structure, and the habits that make any of these models more useful, and it costs nothing. The cheat sheet is there when you want the full reference.