OpenAI introduces GPT-6.2 Astra and Sol

OpenAI announced GPT-6.2 Astra and Sol on November 17, describing improvements in professional work, coding, computer use and efficiency across ChatGPT Work, Codex and the API. The company says both models meet its Critical cybersecurity capability threshold and High biological and chemical threshold. Astra additionally meets its High AI self-improvement threshold.
The company reports stronger performance on exploit-development tasks than its earlier models and says it has tightened controls on higher-risk cybersecurity assistance. These results come from OpenAI's own evaluations and need independent corroboration before being treated as comparative performance facts.
New model capabilities create both developer opportunities and security concerns.
Read full briefing Collapse briefing
Detailed briefing
Model positioning
Astra and Sol represent different capability and efficiency choices for professional tasks and coding agents. Meaningful deployment comparisons should measure successful repository-level tasks, latency, supervision and total cost rather than relying on benchmarks alone.
Cybersecurity assessment
The published system card reports stronger ability to build exploits from known vulnerabilities. OpenAI says this prompted tighter default restrictions and additional vetting for more permissive access.
What remains unproven
Vendor tests do not establish independent model superiority or actual enterprise reliability. The system card is useful evidence of the company's testing methodology and risk classifications, not a substitute for external evaluation.
Limitations & caveats
Vendor-reported evaluations.
