OpenAI 已悄悄回退 ChatGPT 的「模型路由器」：对 Free 与每月 5 美元的 Go 方案，不再自动把复杂问题导向较慢、更昂贵的推理模型，而是预设使用 GPT-5.2 Instant。Free/Go 仍可用推理模型，但需手动选择。该路由器约 4 个月前推出，原本想用「自动分流」取代令人困惑的模型选单。

实际上，路由器让免费用户使用推理模型的比例从不到 1% 升到 7%，增加了服务成本；消息人士称也拖累每日活跃用户。推理模型可能需要数分钟，消费者更在意回复速度；若要等待约 20 秒的「思考点点」，互动感就会变差。OpenAI 表示依据回馈，Free/Go 更偏好维持预设体验，再视需要手动切到推理模式。

竞争压力也在上升：ChatGPT 拥有超过 8 亿每周活跃用户，但外部数据显示 Gemini 近月成长明显、ChatGPT 增长趋平；SimilarWeb 指出自 9 月起，ChatGPT 的平均停留时间低于 Gemini。安全面上，OpenAI 提到每周有「数十万」用户出现可能的精神健康紧急迹象；它称因 GPT-5.2 Instant 在安全基准上表现提升，将不再把敏感对话自动路由到推理模型。

OpenAI quietly rolled back ChatGPT’s “model router” for the Free tier and the $5/month Go plan. Instead of automatically sending harder prompts to slower, costlier reasoning models, these users now default to GPT‑5.2 Instant. Reasoning models remain available, but only via manual selection. The router launched about four months earlier as part of a push to simplify the experience and reduce reliance on a confusing model picker.

In practice, the router drove far more free users into reasoning models: Sam Altman said usage rose from under 1% to 7%, increasing serving costs. A source told WIRED the change also hurt daily active users. Because reasoning models can take minutes, most consumers prioritize speed; even ~20 seconds of “thinking dots” can reduce engagement. OpenAI said feedback showed Free/Go users prefer the default chat flow, opting into reasoning only when needed.

The rollback lands amid intensifying competition. ChatGPT still has 800+ million weekly active users, but third‑party data suggests Google Gemini has been growing while ChatGPT’s growth has flattened; SimilarWeb says ChatGPT’s average visit duration has been below Gemini since September. On safety, OpenAI reported “hundreds of thousands” of users show possible weekly mental‑health emergency signals and said GPT‑5.2 Instant’s improved safety benchmark performance means sensitive chats will no longer be routed to reasoning models.