You start an AI agent on a two-hour job, then close the laptop. When you come back, it has touched a file you never mentioned.
That kind of drift is exactly what ChatGPT Astra was built around. OpenAI made it the headline argument for its new model, ahead of any score.
GPT-6 Astra is the model that replaces GPT-5.6 Sol at OpenAI. It drives a computer and a browser on your behalf, writes documents and code, and ships inside ChatGPT Plus, Pro, Business and Enterprise. In the API it answers to gpt-6-astra.
Here is what it does better than its predecessor, what its numbers are actually worth, where it loses, and why a perfectly legitimate task can end up on hold.
What is GPT-6 Astra?
Astra is a language model, like the one that answers you in ChatGPT, but trained to act rather than chat. It clicks, fills in fields, reads what appears on screen and chains steps on its own.
Quick bearings inside the lineup: GPT-5.6 Sol was the previous flagship, GPT-6 Astra takes its place. Pro, Business and Enterprise plans also unlock a variant called GPT-6 Astra Pro. In the API, on Microsoft Azure and on AWS Bedrock, the model is named gpt-6-astra.
OpenAI lists in its GPT-6 Astra announcement a set of jobs the model can carry from start to finish:
- Data entry: filling online forms, updating customer records in a CRM, tidying up a calendar.
- Research: browsing the web and writing the summary straight into your inbox or your word processor.
- Analysis: opening scientific data, building charts, checking a website page by page.
- Troubleshooting: installing software, testing it, making sense of an error message on screen.
One detail matters if you write code. Inside Codex, OpenAI's development tool, Astra carries notes from one context window to the next instead of squeezing everything into a summary. The context window is the amount of text a model can hold in mind at once.
Earlier exchanges stay readable, so an instruction you gave three hours ago no longer vanishes. OpenAI calls this option experimental, switched on in the config.toml file, and plans to make it the default for Astra.
Driving an agent like this on your own projects is a skill of its own, and our Claude Code course covers the same prompting and context habits.
The numbers, and how to read them
Every result below comes from OpenAI's own tests, run in its research environment.
| Test | What it measures | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|
| OSWorld 2.0 | Real tasks on a computer | 72.6% | 65.7% |
| ScreenSpot-Pro | Finding the right element on screen | 92.7% | 76.9% |
| Terminal-Bench 4.0 | Command-line work | 57.9% | 37.3% |
| FrontierMath Tier 4 (v2) | Unseen mathematics problems | 97.6% | 83.0% |
| ARC-AGI-3 | Abstract reasoning | 99.9% | 7.8% |
| MRCR v2, 512K to 1M | Finding a fact buried in a very long text | 96.3% | 73.8% |
| Internal hallucination test | False claims, lower is better | 4.2% | 12.2% |
The gain you feel most day to day is not a score, it is a clock. On the OSWorld simulations, Astra reaches its 72.6% in roughly 40 minutes per task, against 75 minutes for Sol at 65.7%.
That is 47% less time for a better result. Add the Codex update on top, and OpenAI measures jobs finished 1.9 times faster than with Sol on the Mind2Web test.
The FrontierMath figure deserves a sentence of its own. This test collects maths puzzles hard enough to keep professional researchers busy for hours, and it was designed to sit out of reach of models.
A model that ticks 97.6% of the boxes almost never fails on those puzzles. The test has become too easy to track progress any further, and a harder one will have to be written.
Where Astra is not in front
OpenAI calls Astra "the smartest model in the world". Its own tables tell a more mixed story as soon as you look at the indexes published by third parties.
- Artificial Analysis Intelligence Index v4.1.1 runs each model through about ten different tests and averages them into a single general-level score: 61.2 for Astra, against 65.7 for Claude Fable 5.1 and 63.1 for Claude Opus 5.
- Humanity's Last Exam with tools asks questions written by researchers inside their own field, in medicine or physics for instance, with web search allowed during the exam: 57.2% for Astra, against 65.0% for Claude Fable 5.1.
- Artificial Analysis Coding Agent Index v1.4 measures coding alone across several files, with no developer fixing things between steps: 67.0 for Astra, 67.2 for Claude Fable 5.1 and 68.1 for Claude Fable 5.
Astra's lead sits exactly where OpenAI aimed: driving a screen, holding a long context, solving maths, producing exploits. On expert general knowledge and on coding agents taken as a whole, Anthropic's models match it or move ahead.
The first model rated critical for cybersecurity
OpenAI grades its models against an internal preparedness framework, which weighs the risk of serious harm before each release. Astra is the first one the company places at the "critical" level for cybersecurity, the highest rung.
In practice, it finds unknown flaws and builds working attacks against well-defended systems, without a human steering every step.
The measurements OpenAI published in its safety update on Astra and in the model announcement:
- ExploitBench, turning a known flaw into a working attack: 100% for Astra, 78.5% for GPT-5.6 Sol.
- ExploitBench June to August 2026 edition, built on recent flaws so the model cannot have met them in training: 39.0% against 11.5% for Sol.
- SRE-Bench, working out how a piece of software behaves without its source code: 88.0% on the first try, 99.2% within four attempts, against 55.9% and 68.7% for Sol.
Outside experts hired for the job also watched the model break out of the sandbox of a hardened browser, the walled-off space where testers run their experiments, run commands on the host machine, then gain administrator rights on an operating system.
Those results describe the model tested without its production safeguards, and with the wider access granted by the Daybreak programme. The version you get inside ChatGPT will review code and write patches, but turns down more offensive requests, including building a proof-of-concept exploit. Do not count on it for a penetration test: what you get is a deliberately restrained model.
Alignment, and the monitor that cuts in
An "aligned" model is one that does what it was asked, and nothing beyond that.
After the breach at Hugging Face, the platform where open models are shared, OpenAI built a dedicated test.
The idea: hand the model a hard or impossible task, and watch whether it steps outside its remit to get through. Without production safeguards, GPT-5.6 Sol overstepped in 48% of cases, Astra in 0%.
OpenAI warns that these checks sometimes slow down, pause or kill work that is perfectly legitimate, including work with nothing to do with security. Long-running agents are hit first. Inside ChatGPT or Codex, you are asked to approve before things continue. In the API, the task stops.
Pricing, access and data
Astra first went out to a limited set of organisations, then reached ChatGPT Plus, Pro, Business and Enterprise, plus the API, Microsoft Azure and AWS Bedrock. Free accounts are not part of the plan, the same accounts that saw advertising land in ChatGPT in France.
For developers, OpenAI announced at launch $10 per million input tokens and $50 per million output tokens, with separate rates for cache. A fast mode doubles both the speed and the price. The OpenAI API pricing page holds the figures of the day.
A sense of scale: one million input tokens plus one million output tokens comes to $60. A million tokens is roughly 750,000 words, eight to ten thick novels. Output costs five times what input costs, so it is the length of the answers that drives your bill, not the length of your documents.
Astra usage counts against the quotas already bundled in each subscription, and OpenAI sells credits for going further. On Enterprise accounts, access was switched off by default at launch: an administrator has to turn it on for the workspace.
Eligible API customers can ask for zero data retention, which keeps exchanges from being stored. OpenAI says it is also testing a private safety pipeline, meant to keep the monitoring in place without exposing customer content.
Frequently asked questions
How do you get access to ChatGPT Astra?
GPT-6 Astra is available on the ChatGPT Plus, Pro, Business and Enterprise plans, as well as through the OpenAI API under the name gpt-6-astra, on Microsoft Azure and on AWS Bedrock.
On Enterprise workspaces, access was off by default at launch and had to be enabled by an administrator. Pro, Business and Enterprise also include the GPT-6 Astra Pro variant.
How much does GPT-6 Astra cost in the API?
OpenAI announced at launch $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and cache writes.
A fast mode exists, up to twice as quick for twice the standard price. Current amounts are listed on the OpenAI API pricing page.
Does Astra beat Claude on every test?
No. On the Artificial Analysis Intelligence Index v4.1.1, an index where the highest score wins, Claude Fable 5.1 scores 65.7 against 61.2 for Astra.
The gap repeats on Humanity's Last Exam with tools, 65.0% for Claude Fable 5.1 against 57.2% for Astra. Astra does lead on computer use, mathematics and cybersecurity.
Why does my task stop halfway through?
OpenAI runs a monitor in production that inspects the model's reasoning and actions, and automatically halts anything it judges not allowed, including by mistake.
In ChatGPT and Codex, you are asked to approve the action before work resumes. In the API, the task stops with no way to pick it up again.
Can Astra help with defensive cybersecurity?
Yes for secure code review and writing patches, which OpenAI names explicitly as permitted uses of the public version.
No for more offensive work, such as building a proof-of-concept exploit for a flaw: the model refuses. OpenAI keeps those uses behind the wider access of its Daybreak programme.