OpenAI released GPT-6 Astra in a limited preview on September 3, 2026, then widened the rollout the next day to ChatGPT Plus, Pro, Business, and Enterprise users, plus the OpenAI API and AWS. Astra is the first OpenAI model to hit the “Critical” rating for cybersecurity capability under the company’s Preparedness Framework, the internal scale OpenAI uses to grade how dangerous a model’s skills could be if misused. That rating comes with real consequences: the model’s most advanced cybersecurity abilities are switched off by default, even for paying customers, and enterprise admins have to explicitly turn them on.
What GPT-6 Astra actually does
According to OpenAI’s own announcement, Astra saturates FrontierMath Tier 4 at a 98% score and ARC-AGI-3 at 99.9%, and it hits a perfect 100% on ExploitBench, a benchmark built around finding and chaining software vulnerabilities. OpenAI VP of Research Aidan Clark said the model came out of the company’s largest training run yet, the first time OpenAI pretrained a model on more than 100,000 GPUs, at its Stargate site in Texas. Company president Greg Brockman went further, saying Astra could eventually be seen as the arrival of artificial general intelligence.
Computer use and general task speed both got a real bump too, with OpenAI describing Astra as noticeably faster at browsing, coding, and multi-step work than the models that came before it, including the GPT-5.6 Sol Ultrafast mode the company shipped in August.
Why the Critical rating actually matters
A perfect ExploitBench score is exactly why OpenAI didn’t just ship Astra wide open. Under the Preparedness Framework, a Critical cybersecurity rating means the model is treated as capable enough to meaningfully assist in real attacks, not just answer trivia about them. OpenAI is routing the model’s most sensitive capabilities through a program called Daybreak, with initial access limited to vetted testers and a defensive-only track called Daybreak Blue for security teams. For everyone else, including Enterprise customers, the advanced cybersecurity behavior is simply off unless an admin flips it on.
This is a notably different approach from how Google handled its own cyber-capable release earlier this month. The Gemini 3.8 Flash cyber variant shipped as a separate, purpose-built model aimed at defenders. Astra is OpenAI’s flagship general model that happens to have crossed a threshold nobody expected this soon, which is part of why the gating decision made news on its own.
What this means if you build on OpenAI’s models
Pricing-wise, Astra usage is folded into existing ChatGPT and API subscription allowances, with extra credits available for heavier use. The bigger thing to plan around is the access model itself. If a product integrates OpenAI’s API for anything security-adjacent, from a code review assistant to an internal vulnerability scanner, expect that integration to need explicit enablement going forward rather than working the moment a new model ships.
That’s a good prompt to check how a client’s stack actually depends on a given model version and what happens when a provider changes what’s available by default. Teams running production AI features through ConicPlex’s AI Development work already build in that kind of review, since a capability gate like this one can quietly change what an integration is allowed to do overnight.
