OpenAI GPT-6.2 arrived on November 17, 2026 with two models, Astra and Sol. Both are rolling out in ChatGPT Work, Codex and the API as gpt-6.2-astra and gpt-6.2-sol. The release focuses on lower prices and better efficiency, as OpenAI frames it.
Astra: cheaper and sharper
Astra API pricing is $8 per million input tokens and $40 per million output tokens. That is 20 percent below GPT-6 Astra. In addition, OpenAI says Astra gains 6.2 points on GDPval v2 while using 22 percent fewer output tokens. As a result, teams may pay less for better answers.
Sol: near-Astra quality for less
Sol costs $2 per million input tokens and $10 per million output tokens, unchanged from GPT-6.1 Sol. However, it comes within 0.5 points of Astra on FrontierCode 2.0. It does so at roughly a fifth of the cost. For coding tools and agents, that trade-off could matter a lot.
Safety flags to know about
Both models meet the Critical threshold for cybersecurity and the High threshold for bio and chemical risk under OpenAI’s own framework. Moreover, Astra is the first externally released model to meet the High threshold for AI self-improvement. Those labels do not mean harm is happening. Still, they show the guardrails are getting heavier.
How it ranks against rivals
Independent tests cited by Decrypt rank Astra second, behind Claude Fable 5.2. Benchmarks vary by task, so results may shift. Meanwhile, the price cut puts pressure on other labs to respond. Watch API bills over the next month to see who moves first.
If you build on OpenAI, test both models on your own workload. Sol may be enough for most tasks, while Astra is the pick for harder reasoning. Also, review the new safety thresholds if your product touches security or science.
What should developers do now? Switch the model name in the API and compare cost and quality side by side on your own prompts. Because pricing is per million tokens, even small efficiency gains add up at scale. Also, ChatGPT Work and Codex get the models too, so the rollout is not limited to API customers. In the end, real workloads will tell you more than any leaderboard.
