GPT-6 Astra: What Actually Changed

Read the benchmark table carefully

Every number in the table above comes from OpenAI's own published comparison, including its runs of Anthropic's models. That is normal industry practice and the figures are disclosed with methodology footnotes, but a vendor's scoring of its competitor is not the same thing as an independent test. Anthropic's own published table for Fable 5.1 uses different evaluations and, by the same logic, flatters Anthropic.

What survives that caution is the shape of the result rather than the decimal places. Astra clearly leads on agentic coding, mathematics and abstract reasoning. Fable 5.1 leads on Humanity's Last Exam with tools (65.0% vs 57.2%) and on the third-party Artificial Analysis Intelligence Index (65.7 vs 61.2) — the one column in OpenAI's own table that OpenAI does not win. The two models are genuinely close, and which one is "better" depends entirely on the task.

Computer use is the real change

The headline capability is not chat. Astra is built to operate a computer — filling in forms, updating records in a CRM, running browser-based research, driving specialist software, and checking its own work in a browser afterwards.

OpenAI reports 72.6% on OSWorld 2.0 against 65.7% for GPT-5.6 Sol, achieved in roughly 47% less time per task, and 92.7% against 76.9% on ScreenSpot-Pro. Paired with an updated Codex harness, OpenAI claims 1.9x faster task completion than the previous generation on Mind2Web.

If you have tried delegating multi-step desktop work to an assistant and given up because it was too slow or too unreliable, this is the release that most directly targets that complaint. If you mostly use ChatGPT to write and answer questions, the practical difference will be far smaller than the launch coverage implies.

The monitorability tradeoff OpenAI disclosed

Astra uses a technique that lets it solve problems with fewer written reasoning steps. A side effect is that there is less written reasoning to inspect. In OpenAI's words, its evaluations found Astra's reasoning harder to monitor than GPT-5.6 Sol's on tests that explicitly asked the model to evade monitoring, and OpenAI says it takes the decline seriously.

This cuts both ways and it is worth being precise about which way. On the alignment measures OpenAI published, Astra is better behaved: on an internal computer-use safety benchmark it produced misaligned outcomes 2.4% of the time against 22.0% for Sol, and on a test built after the Hugging Face incident it went beyond its authorised scope in 0% of cases against 48% for Sol. The concern is not that Astra misbehaves more — it is that if it did, the written trace would tell you less.

For everyday use this is an abstraction. For anyone running unattended agents on systems that matter, it is a reason to lean on run logs, approval gates and auto-review rather than on reading the model's stated reasoning.

Pricing, and the 272K cliff

Standard API pricing is $10 per million input tokens and $50 per million output. Cached input is $1.00 per million and cache writes are $12.50 per million — cache writes bill at 1.25x the uncached input rate.

The trap is the long-context surcharge. The window is 1,050,000 tokens, but any request whose input exceeds 272,000 tokens is billed at 2x the input and cache rates and 1.5x the output rate for the entire request — not just the tokens past the line. A 280,000-token prompt therefore costs meaningfully more than a 270,000-token one. If you are near that boundary, trimming below it is the single cheapest optimisation available.

Two levers go the other way: Batch and Flex processing are priced at 50% of Standard, and Fast mode costs 2x for up to double the speed. On a ChatGPT subscription none of this applies directly — Astra usage comes out of your existing plan allowance, with the option to buy credits for more.

Can you actually get it right now?

Partly. Astra rolled out from September 3 across Plus, Pro, Business and Enterprise, plus the API, Azure and Bedrock. Then on September 10, OpenAI product leader Tibo Sottiaux announced that new sign-ups to the $200/month Pro plan were being disabled because Pro puts the heaviest strain on OpenAI's infrastructure and demand was, in his words, unprecedented.

Existing Pro subscribers keep their access. The Go, Plus, Business and Enterprise plans and the API all remain open. OpenAI has not said when Pro sign-ups reopen.

One more wrinkle for teams: on Enterprise, Astra is off by default at launch and an administrator has to enable it for the workspace. If you are on a company plan and cannot find Astra, that is the likely reason rather than a rollout delay.

Astra or Claude Fable 5.1?

They cost the same per token and they are closer than either vendor's marketing suggests. The honest split, based on what each company published:

Astra is the better pick for driving a computer or browser, long agentic coding runs, competition-grade mathematics, and work that needs more than a million tokens of context in one request. It is also the stronger choice if you want a single model for terminal work and desktop automation.

Fable 5.1 is the better pick for hard research-style reasoning — it leads on Humanity's Last Exam with tools and on the independent Artificial Analysis index — and its 75% cache-read price cut makes heavily agentic API workloads cheaper in practice than the matching headline rates imply.

And the answer most readers actually need: if you are on a subscription and mostly write, research and ask questions, neither flagship is the right daily driver. Both burn allowance fast. The cheaper mid-tier model in each lineup does that work without the cost, which is the same conclusion we reach in Claude Opus vs Sonnet. Save the flagship for the tasks that genuinely escalate past it.

Common questions

Related

• Claude Fable 5.1: what changed
• Claude Opus 5.5 and the same-day GPT-6 Sol launch
• Claude vs ChatGPT
• Best AI for coding