GPT-6 Astra: OpenAI’s new model, what it can do, and who gets it first

OpenAI has released GPT-6 Astra, a model which it says is now the most advanced LLM in the world.

In a blog post on Friday (4 September), OpenAI it as its most intelligent and most aligned model to date. The company said Astra leads the field on computer use, browsing, software engineering, cybersecurity, science and professional work.

The rollout starts with a limited set of organisations before the wider ChatGPT release. Developers will get the model through the OpenAI API as gpt-6-astra, as well as through Microsoft Azure and Amazon Bedrock.

Astra usage counts against existing subscription allowances, and users and businesses can buy credits for additional usage. Pro, Business and Enterprise subscribers also get access to GPT-6 Astra Pro.

Enterprise administrators need to switch the model on for their workspace, as access is off by default at launch.

API pricing is $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and writes. A Fast mode runs at up to twice the speed of Standard processing at twice the Standard price.

What the model can do

OpenAI said Astra can take on everyday computer tasks, including:

• Filling out online forms
• Updating customer records in a CRM
• Organising a calendar
• Researching online and drafting summaries in email or a document editor
• Analysing scientific data and generating plots
• Creating a website and running frontend checks to confirm its features work
• Installing and testing software and troubleshooting problems on screen

OpenAI showed the model laying out a printed circuit board in KiCad, modelling a house in Blender and turning it into a walkable Unreal Engine 5 scene, and completing a US Form 1040 tax return. It also showed Astra searching for a paediatrician, hunting for an apartment and booking a DMV appointment, with one demo running two minutes and 54 seconds.

On Terminal-Bench 4.0, which tests agents on terminal tasks such as software engineering and system configuration, Astra scored 57.9% against 55.8% for Anthropic’s Claude Fable 5.1 and 37.3% for GPT-5.6 Sol.

OpenAI said Astra reached that score at around 63% lower estimated API cost per task than Fable 5.1.

On the OSWorld 2.0 computer-use test, Astra scored 72.6% at roughly 40 minutes per task, compared with 65.7% at roughly 75 minutes for GPT-5.6 Sol. On Agents’ Last Exam, which covers professional tasks such as financial modelling and media production, Astra scored 59.3% against 55.5% for Claude Opus 5.

“Not only is this the best model we’ve ever tested, but it also represents a meaningful step change in frontier-model performance – not only in its ability to navigate and solve novel environments, but also in how efficiently it learns to do so,” said Greg Kamradt of the ARC Prize Foundation.

OpenAI also updated its Codex harness alongside the launch, which it said makes task completion 1.9 times faster than the current GPT-5.6 Sol experience on the Mind2Web benchmark.

Astra in Codex can keep notes across context windows instead of compressing earlier work into a single summary, and users can enable this experimental feature in the Codex config.toml file.

Cybersecurity limits

OpenAI said Astra meets the Critical threshold for cybersecurity under its Preparedness Framework.

Without production safeguards, the model scored 100% on ExploitBench, which tests whether a model can turn known software vulnerabilities into working exploits, up from 78.5% for GPT-5.6 Sol.

Astra discovered and used two previously unknown zero-day vulnerabilities during testing, and OpenAI said it is disclosing both to their maintainers.

Expert-led assessments found the unrestricted model could achieve arbitrary code execution in hardened browsers and build privilege-escalation exploits for hardened operating systems.

The version available at launch refuses more advanced tasks such as creating proof-of-concept exploits for vulnerabilities. OpenAI said it plans to roll out less restrictive safeguards through OpenAI Daybreak in the coming weeks for defensive work including malware analysis and detection engineering.

Extra safety checks may slow, pause or stop legitimate work, including defensive cybersecurity tasks. When a task pauses in ChatGPT or Codex, users may need to review the action before continuing, while in the API the task stops.

In OpenAI’s computer-use safety stress test, Astra produced misaligned outcomes in 2.4% of cases, against 9.5% for Claude Fable 5.1 and 11.5% for Claude Opus 5. OpenAI also said Astra’s written reasoning is harder to monitor than GPT-5.6 Sol’s, and that improving monitorability remains a research priority.

Now read: FCA warns frontier AI is finding cyber vulnerabilities faster than firms can fix them