OpenAI says its newest model, GPT-6 Astra, is its most capable yet — and its riskiest
OpenAI released GPT-6 Astra, which it calls its most intelligent and aligned model to date, with new capabilities in computer use, coding, cybersecurity, and science. By OpenAI's own safety framework, it's also the first model the company has rated at the highest ("Critical") level of cybersecurity risk.
Open in the XNewsAi app →Why this matters
For the 806
Both of those claims — "most capable" and "Critical cybersecurity risk" — come from OpenAI's own testing, not an outside lab. No independent group has confirmed either one yet. Worth watching what independent security researchers say once they get access.
What we know
- OpenAI released GPT-6 Astra on September 3, 2026, describing it as its most intelligent and aligned model yet, with new capabilities in computer use, coding, cybersecurity, and science.
- OpenAI's own Preparedness Framework rates GPT-6 Astra at the "Critical" level of cybersecurity capability — the first model the company has rated that high.
What we don't know
- No independent (non-OpenAI) lab or researcher has yet confirmed either the capability claims or the Critical cybersecurity rating.
- The exact date the model finished training isn't disclosed — this is the announcement date.
Sources
- PrimaryOpenAI — GPT-6 Astra announcement ↗
OpenAI's own release announcement.
- PrimaryOpenAI — Safety overview: GPT-6 Astra ↗
OpenAI's own safety-framework assessment of the model.