Skip to content
Advertisement
OpenAI says GPT-6.2 Astra is first deployed model to hit High self-improvement threshold

OpenAI says GPT-6.2 Astra is first deployed model to hit High self-improvement threshold

FoxTPNL / Wikimedia Commons (CC BY 4.0)

OpenAI's new flagship crossed a line in its own Preparedness Framework, while sibling model GPT-6.2 Sol landed at Critical for cybersecurity

OpenAI has released a model that, by its own measurement, is meaningfully good at helping build better AI.

GPT-6.2 Astra is the first externally deployed model to meet the High threshold for AI Self-Improvement under OpenAI’s Preparedness Framework. The company built that scorecard to track exactly this kind of capability.

The model arrived on November 17, 2026, alongside a sibling called GPT-6.2 Sol. Both are now live in ChatGPT Work, in Codex and through the API.

Two models, two very different risk report cards

OpenAI is positioning the pair as a split lineup. Astra is the heavier, more capable option. Sol is the cost-effective model meant for everyday use.

Both are built for professional workloads, and both are described as more efficient than what came before.

The safety classifications are where things get interesting. OpenAI’s Preparedness Framework sorts dangerous capabilities into categories and grades each one against tiers. The labels named in this release are High and Critical.

Advertisement

Astra is the headliner because it cleared the High bar for self-improvement. No externally deployed OpenAI model had done that before.

Sol did not reach High on self-improvement. Sol meets the Critical threshold for cybersecurity and the High threshold for biological and chemical risks.

How Astra stacks up against its predecessor

Astra replaces GPT-6 Astra, which OpenAI released in September 2026. That earlier model posted a perfect score on ExploitBench. Even so, it fell short of the High self-improvement threshold.

Roughly two months later, the successor clears that bar. On the benchmarking side, GPT-6.2 Astra improves on GPT-6 Astra by 5.2 points on Agents’ Last Exam.

OpenAI also cut the price. GPT-6.2 Astra costs less than GPT-6 Astra, with general-user pricing set at $8/$40 per million tokens.

Why the self-improvement category matters

Most AI risk categories describe what a model could help a person do. Think of assisting with a cyberattack or providing dangerous chemistry knowledge.

Self-improvement is a different kind of concern. It asks whether a model can contribute to the development of AI itself, potentially speeding up its own successors.

That is the idea behind the phrase recursive self-improvement. OpenAI says it is continuing to evaluate that capability as part of its ongoing safety work.

What this means for developers, businesses and the industry

For developers and enterprise buyers, the immediate story is practical. A more capable flagship is available at a lower price than its predecessor. A cheaper everyday option ships alongside it.

That two-tier structure lets teams route demanding work to Astra and routine tasks to Sol.

Distribution matters too. Shipping both models through ChatGPT Work, Codex and the API at once puts them in front of business users, coders and developers on day one.

The safety picture is harder to wave through. OpenAI is deploying a model at High for self-improvement and a model at Critical for cybersecurity in the same release.

Those classifications come from OpenAI’s own framework. That makes them a self-reported warning label, not an outside audit.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.