发生了什么
OpenAI 表示,在 Responses API 中使用 GPT-6.2 的智能体可以基于先前的推理继续推进,从而以更少的 token 和更高的缓存命中率交付更好的结果。从其最新一代前沿模型开始,每个推理项都会与之前的对话历史绑定,因此编辑较早的消息会导致 API 丢弃后续的推理。
推荐理由
当较早的消息被编辑时,Responses API 会丢弃与此前对话历史绑定的后续推理项,而使用 GPT-6.2 的智能体则可以基于先前的推理继续推进,同时减少 token 消耗并提高缓存命中率。
Agents using GPT-6.2 in the Responses API can build on earlier reasoning to deliver better results with fewer tokens and more cache hits.
Starting with our newest frontier models, each reasoning item is now tied to the conversation history before it, so editing earlier messages causes the API to drop reasoning that follows the change.
2 篇报道
这件事到目前为止
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol in ChatGPT Work, Codex, and the API, with Astra priced 20% lower per input and output token than GPT-6 Astra and Sol keeping GPT-6.1 Sol's token pricing. OpenAI also expanded Ultrafast mode to Pro 500 subscribers for both models and to Pro 200 subscribers starting with Sol, while cutting Ultrafast API and Enterprise pricing by 50% for three months. Because of increased capabilities, OpenAI launched both models with tighter cybersecurity safeguards that it said may increase refusals on some security research tasks, and advised developers using the Responses API to append new turns rather than edit earlier messages, as editing drops later reasoning tied to the conversation history. OpenAI reported new highs for Astra on Agents' Last Exam, BenchCAD 2.1 Core, and Terminal-Bench Science 0.1, while Greg Brockman said Astra exceeded the human baseline on a physical 3D understanding benchmark, and Perplexity reported higher scores on its WANDR benchmark after adding both models to Perplexity Search and Computer. Independent feedback has been mixed: Epoch AI reported that Astra scored 65% on its Epoch Automation Reports benchmark, trailing Fable 5.2's 82%, and developer Simon Willison said neither model handled his animation benchmark well compared with Claude Fable 5.2 and Gemini 4 Carbon. Solveit's Jeremy Howard also said he reverted Astra and Sol to pre-6.2 versions because the update was not working for him or his users.
这件事的进展 · 5
查看完整事件- 11月20日 08:00
OpenAI 在 Responses API 中将推理项与对话历史绑定 本条
- 11月19日 00:33Greg Brockman 表示 GPT-6.2 在浏览器使用方面的表现令人惊叹Greg Brockman · X
- 11月18日 06:15Tibor Blaho 在日常工作中测试 GPT-6.2 Astra 和 SolTibor Blaho · X
- 11月18日 06:15Perplexity 分享了 GPT-6.2 Astra 和 Sol 的 WANDR 评估结果Aravind Srinivas · X
- 11月18日 02:36OpenAI 推出 GPT-6.2 Astra 和 SolOpenAI News

