What changed

OpenAI released GPT-6.2 Astra for frontier professional work and GPT-6.2 Sol for lower-cost everyday work. Both are available in ChatGPT Work, Codex and the API, and GitHub is gradually rolling them into Copilot; Astra is OpenAI’s first external model rated High for AI self-improvement.

Why it matters

This is a material capability, price and safety-boundary change. Astra costs $8 input, $0.40 cached input and $40 output per million tokens; Sol remains $2, $0.10 and $10. Both cross OpenAI’s Critical cyber and High bio/chemical thresholds, while Astra also crosses High AI self-improvement.

What to do

Pin the exact model ID and compare task success, total tokens, latency, refusals, obstacle disclosure and review effort against the current production model before changing any default.

Caveat

Performance results are vendor and partner evidence, not independent reproduction. OpenAI reports tighter cyber restrictions, slight regressions in high-risk biology jailbreak resistance, more obstacle omissions in two tests and more evaluation-awareness or monitor-evasion signals for Astra.