- GPT0%
- Token+22.16%
动察 Beating AI 快讯，OpenAI 给 GPT-6 Astra 新增 configuration_update。在 Responses API 的同一段长对话里，现在可以随时调高或调低 reasoning effort，又不改写已经缓存的 Prompt 前缀。难题切 High、xhigh，简单任务再降回 Low，前面的长上下文还能继续复用。
以前这是个很隐蔽的烧钱坑。reasoning effort 会作为上下文开头的指令，中途切到新的思考档位，会导致大量缓存无法继续复用。有 Codex 用户实测，切换档位后缓存命中率曾从 98.5% 掉到 65%，下一轮重新建立缓存后才恢复到 97% 以上。
对长时间运行的 Agent 来说，这个改动很实用。碰到难题可以临时拉高思考强度，简单任务再降回来，不会只因为换挡，就让前面几十万 Token 的缓存重新按普通输入处理。
Disclaimer: The current content is sourced from third-party perspectives or directly translated by AI from third-party perspectives. CoinEx does not guarantee the authenticity, accuracy, and originality of the content, and it does not constitute any investment advice from CoinEx. The prices of cryptocurrencies are highly volatile, please be aware of the potential risks.
- CoinsPrice24H Change