On September 3, 2026, OpenAI released GPT-6 Astra — its most capable model yet, and the first to cross a safety line the company itself calls “Critical.” If you’ve heard the name floating around but aren’t sure what actually changed, here’s the plain-English breakdown.
What Astra Actually Is
GPT-6 Astra is OpenAI’s new flagship model, built for long, complex, multi-step work across coding, browsing, computer use, professional analysis, science, and mathematics. In other words, it’s not just a chatbot that answers questions — it’s designed to do things on your behalf: fill out forms, update spreadsheets, write and test code, and carry out research across many steps without constant hand-holding.
OpenAI says the model brings major improvements in computer use, coding, scientific research, and professional work, letting people automate complex digital tasks more efficiently. The company positions Astra as state-of-the-art across computer use, browser use, software engineering, cybersecurity, science, and professional work.
The Headline Numbers
For the benchmark-curious, a few results stand out:
- On the OSWorld 2.0 computer-use test, Astra scored roughly 73%, up from about 66% for the previous model, GPT-5.6 Sol.
- The model supports context windows of around one million tokens, with strong scores on long-context retrieval tests even near the top of that range.
- Independent evaluators note Astra roughly ties GPT-5.6 Sol on general intelligence benchmarks, while still trailing Anthropic’s Claude Fable 5.1 overall.
In short: it’s a meaningful step up in doing work — using a computer, writing code, running multi-step tasks — even if it isn’t a dramatic leap in raw general intelligence.
The Part That Slowed the Launch: Cybersecurity
This is the detail that makes Astra different from a routine model update. Internal testing showed Astra crossing OpenAI’s “Critical” cybersecurity threshold — meaning it can find and exploit software vulnerabilities largely on its own. That result was serious enough that OpenAI delayed the release in August to add extra safeguards before shipping.
Because of this, access isn’t a simple flip of a switch. At launch, Astra helps with defensive tasks like secure code review and patching, but it declines more advanced requests, such as building proof-of-concept exploits. Looser restrictions are reserved for vetted organizations in OpenAI’s trusted-access program.
For everyday users, this mostly matters as context: it’s a sign that frontier AI is now capable enough that labs are actively gating certain features rather than shipping everything to everyone at once.
Who Gets It, and When
Rollout is staged rather than instant:
- Access starts with organizations enrolled in OpenAI’s application-based cybersecurity program.
- From there, it opens up to ChatGPT Plus, Pro, Business, and Enterprise plans, plus the OpenAI API and Amazon Web Services.
- For business accounts, Astra is off by default — an administrator has to turn it on for their workspace.
If you’re a paying ChatGPT user and don’t see it yet, it’s likely just a matter of days, not a sign anything is wrong on your end.
Should You Actually Care?
If you’re a casual user asking ChatGPT questions day-to-day, Astra’s biggest upgrades — computer use, long-running coding tasks, cybersecurity work — won’t change much of what you see immediately. The real shift is for:
- Developers and technical teams, who get a genuinely stronger coding and agentic-task model, with computer-use tasks now running noticeably faster than before.
- Business users, who can hand off tedious multi-step work like document creation, data organization, and research.
- Security teams, who now have to think seriously about AI models capable of finding vulnerabilities — for both defense and risk.
The Bottom Line
GPT-6 Astra isn’t just a version of ChatGPT that chats better — it’s a model built to act, not just answer. That’s a genuinely useful shift if you want AI to handle real workflows instead of just giving you text to copy and paste. It’s also one of the clearest signs yet that AI capability and AI risk are now advancing together, which is exactly why OpenAI gated the model’s sharpest edges before letting everyone in.
If you’re deciding whether to try it: start with a task you already do manually and repetitively — data entry, research summaries, or routine coding — and see how much of it Astra can take off your plate.
Have questions about GPT-6 Astra, or want a side-by-side comparison with Claude or Gemini? Let us know what you’d like covered next.