Published 4 September 2026
OpenAI released GPT-6 Astra on 3 September 2026 at $10 per million input tokens and $50 per million output tokens. Astra is built to operate software the way a person does, and OpenAI reports 72.6% on an offline subset of OSWorld 2.0 against 65.7% for GPT-5.6 Sol. Access starts with a small group of enterprise customers, because Astra is the first OpenAI model to reach the Critical cybersecurity level under the company’s Preparedness Framework.
What can GPT-6 Astra do?
GPT-6 Astra works through screens and controls instead of needing an API integration for every application. OpenAI shows it filling in online forms, updating CRM records, editing spreadsheets, running web research and driving engineering software such as KiCad and FreeCAD. The pitch is that it finishes the job rather than telling you how to do it.
Speed is the other half of the claim. On OSWorld 2.0, OpenAI reports Astra taking about 40 minutes per task where GPT-5.6 Sol takes about 75 minutes, roughly 47% less time. OpenAI also updated the Codex harness, which it says makes task completion 1.9 times faster than the current Sol experience on the Mind2Web benchmark.
Long sessions get a new trick. In Codex, Astra can keep notes across context windows instead of compressing everything into one summary each time the context fills, and earlier windows stay searchable. The feature is experimental and you switch it on in your Codex config file.
GPT-6 Astra benchmarks and pricing
The table below shows where Astra leads clearly, where it barely leads, and the one independent score where it does not lead at all.
| What is measured | GPT-6 Astra | Claude |
|---|---|---|
| OSWorld 2.0, offline set (operating a desktop) | 72.6% | Claude Opus 5: 70.2% |
| Terminal-Bench 4.0 (terminal and system tasks) | 57.9% | Claude Fable 5.1: 55.8% |
| DeepSWE v1.1 (fixing real software bugs) | 74.1% | Claude Opus 5: 73.7% |
| Artificial Analysis Intelligence Index v4.1.1 (independent overall score) | 61.2 | Claude Fable 5.1: 65.7 |
| ExploitBench (turning known flaws into working exploits) | 100% | GPT-5.6 Sol: 78.5% |
| Price per million tokens, input and output | $10 and $50 | Claude Fable 5.1: $10 and $50 |
Every figure here comes from OpenAI’s own launch page, including the Artificial Analysis scores OpenAI chose to publish. OpenAI ran most of the comparison models itself, and its footnotes say the Claude results on OSWorld and BenchCAD use settings that differ from Anthropic’s own reported runs. Treat the gaps as directional until someone outside OpenAI repeats them.
Fast mode doubles the price to $20 and $100 per million tokens for up to twice the speed. Cache reads and writes are billed separately and OpenAI does not give those rates on the launch page.
How does GPT-6 Astra compare to Claude Fable 5.1 and GPT-5.6 Sol?
Against its own predecessor the gap is wide in exactly the places OpenAI is selling.
- Terminal-Bench 4.0 goes from 37.3% to 57.9%.
- ARC-AGI-3 goes from 7.8% to 99.9%.
- FrontierMath Tier 4 goes from 83.0% to 97.6%.
- Recall on OpenAI’s own long-context test at 512K to 1M tokens goes from 73.8% to 96.3%.
Against Anthropic the picture is much closer.
- On DeepSWE v1.1, the coding benchmark most teams watch, Astra scores 74.1% against 73.7% for Claude Opus 5 and 73.8% for Gemini 3.8 Flash.
- On Humanity’s Last Exam with tools, Astra scores 57.2% while Claude Fable 5.1 scores 65.0%.
- On the Artificial Analysis Intelligence Index, Astra sits at 61.2 against 65.7 for Fable 5.1.
Those are OpenAI’s own published numbers, in OpenAI’s own comparison table. The honest summary is narrower than the AGI framing suggests. Astra is clearly ahead on operating a computer, on cybersecurity and on hard mathematics. On general reasoning and everyday coding it is level with the field.
What OpenAI is not saying about GPT-6 Astra
Two things are missing from the launch page. There is no context window figure: the long-context table runs to 1M tokens, so the window is at least that, but OpenAI never states it as a specification. And there is no general availability date. Astra is rolling out to a limited set of organisations, with paid ChatGPT plans and the API following over the coming days. Enterprise administrators have to switch it on, and it is off by default.
One disclosure that is on the page deserves more attention than it is getting. Astra’s written reasoning is harder to monitor than GPT-5.6 Sol’s. OpenAI attributes this to Astra solving problems in fewer written steps, calls the decline serious, and says monitorability remains a research priority. If your governance plan depends on reading an agent’s chain of thought, read that sentence twice.
What this means
If you run browser or desktop automation, GPT-6 Astra is worth testing as soon as you can get access. That is the genuinely new part: a model that finishes a multi-step job inside real software at roughly half the time per task of its predecessor, and that stayed inside its authorised scope in every one of OpenAI’s honeypot tests where GPT-5.6 Sol strayed 48.2% of the time. For a team that has been hand-writing connectors to every internal tool, that changes the build-versus-drive calculation.
If you run a chat product or a coding agent, this is worth watching rather than switching to. Astra costs the same as Claude Fable 5.1 and sits behind it on both independent indices OpenAI itself published. The upgrade you would be paying for is computer use, and most coding agents do not use one. Add the practical friction: safety checks can slow, pause or stop legitimate work, and in the API a flagged task simply stops. Pilot it on a workload you can afford to have interrupted before you route production traffic to it.
For more information, visit the official announcement of GPT-6 Astra on the OpenAI blog.
