OpenAI 重置 Codex 与 ChatGPT Work 用量限制,详解 GPT-5.6 Sol 效率问题OpenAI Resets Codex & ChatGPT Work Limits, Explains GPT-5.6 Sol Efficiency Issues
OpenAI Codex & ChatGPT 负责人 Thibault Sottiaux 宣布已为所有 ChatGPT Work 和 Codex 用户重置用量限制,并恢复五小时限额。他详细解释 GPT-5.6 Sol 比预期更快消耗额度的原因:模型更愿意长时间工作、进行额外工具调用、协调复杂工作流;在相同推理强度下比前代更努力;程序化工具调用(code mode)带来更多并行响应和缓存 token。团队已优化工具等待与网页搜索处理,预期典型使用可延长约 18%。他承认能力与效率不同步,本应更早识别长尾高消耗用户问题。另有一句:「有一天我们创造了重置按钮,其余皆成历史。」
OpenAI Codex & ChatGPT lead Thibault Sottiaux announced a full reset of usage limits for all ChatGPT Work and Codex users, plus restoration of the five-hour limit. He detailed why GPT-5.6 Sol drained quotas faster than expected: the model is far more willing to work longer, make extra tool calls, and coordinate complex multi-tool workflows; it works harder at the same reasoning effort than prior models; and programmatic tool calling (code mode) produced more responses per turn and higher cached input tokens. Improvements to tool-wait handling and web search should extend typical usage by around 18%. He acknowledged that capability and efficiency do not always advance in lockstep and that the team focused too much on median usage before launch.
OpenAI Codex & ChatGPT lead Thibault Sottiaux announced a full reset of usage limits for all ChatGPT Work and Codex users, plus restoration of the five-hour limit. He detailed why GPT-5.6 Sol drained quotas faster than expected: the model is far more willing to work longer, make extra tool calls, and coordinate complex multi-tool workflows; it works harder at the same reasoning effort than prior models; and programmatic tool calling (code mode) produced more responses per turn and higher cached input tokens. Improvements to tool-wait handling and web search should extend typical usage by around 18%. He acknowledged that capability and efficiency do not always advance in lockstep and that the team focused too much on median usage before launch.
OpenAI Codex & ChatGPT lead Thibault Sottiaux announced a full reset of usage limits for all ChatGPT Work and Codex users, plus restoration of the five-hour limit. He detailed why GPT-5.6 Sol drained quotas faster than expected: the model is far more willing to work longer, make extra tool calls, and coordinate complex multi-tool workflows; it works harder at the same reasoning effort than prior models; and programmatic tool calling (code mode) produced more responses per turn and higher cached input tokens. Improvements to tool-wait handling and web search should extend typical usage by around 18%. He acknowledged that capability and efficiency do not always advance in lockstep and that the team focused too much on median usage before launch. He also noted: "One day we created the reset button and the rest is history."
查看原文 →查看原文 →