OpenAI 發佈 GPT-6.2 Astra 和 Sol,Astra 價格更低且評價不一
OpenAI ships GPT-6.2 Astra and Sol: cheaper tokens, mixed reviews
OpenAI 於 17 November 2026 發佈 GPT-6.2 Astra 和 GPT-6.2 Sol,兩款模型均已可透過 API 使用。OpenAI 稱兩款模型在高要求任務上以較少輸出 tokens 取得更好結果;Astra 定價為 $8/$40 per million tokens,Sol 沿用 GPT-6.1 的 $2/$10 定價。
文章並列 GPT-6.2 Astra 與 Sol 的基準表現、定價和獨立測試回饋,呈現兩款模型在成本、成績與實際使用評價上的差異。
Posted on Nov 18 Edited on Nov 18
📅 17 November 2026 • OpenAI
OpenAI shipped two new models on 17 November 2026: GPT-6.2 Astra, a frontier model for the hardest tasks, and GPT-6.2 Sol, an efficient everyday model. Per OpenAI, both deliver better results on demanding work with fewer output tokens, which in practice means more tasks finished inside the same subscription limits. Both are available via API right now as gpt-6.2-astra and gpt-6.2-sol, and Simon Willison has already shipped LLM 0.38 with support for the pair.
Per OpenAI, Astra scores 5.2 points above GPT-6 Astra on Agents' Last Exam at under a third of the cost, and leads Claude Fable 5.2 by 5.5 points at 5% of its cost. On Terminal-Bench-Science 0.1 and the Vals Science Index the model sets new highs, with OpenAI calling it the best in the world for scientists. Pricing, per OpenAI and confirmed by Simon Willison: $8/$40 per million tokens, 20% cheaper than GPT-6 Astra, cached input at $0.40 after a 60% discount, while Sol keeps GPT-6.1 pricing at $2/$10.
For agent builders, the 60% cache discount is the quiet headline: per Simon Willison, cheap cached input is good news for long agent sessions, where the same context gets paid for over and over. Sol is the other practical story, per OpenAI, many OpenAI employees already use it as their default model for technical and non-technical tasks. On Cognition FrontierCode 2.0 Sol trails Astra by just 0.5 points at roughly a fifth of the cost, and on LifeVal it beats GPT-6 Astra at about a seventh of the cost.
The independent verdict is mixed, per Simon Willison: on different benchmarks GPT-6.2 Astra does not always score as high as Claude Fable 5.2, and users complain about a confusing communication style, with Theo Browne noting the model asks odd questions about internal implementation when working with subagents. In his own animation-prompt test, both 6.2 models trailed Fable 5.2 and Gemini 4 Carbon. Willison also flags the footnote that the Fable 5.2 leaderboard result excludes nine tasks where a fallback model was used.
Sources: openai.com, simonwillison.net, latent.space, the-decoder.com, arstechnica.com
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
來源:DEV Community · dev.to