Are Older ChatGPT Models Being Quietly Nerfed?
Big_Wave9732 · reddit · 2026-07-13
作者表示，自己常用 ChatGPT 处理法律与事实研究、文书草拟和法院文件分析，不依赖编程或 agent 工作流。
在新模型即将发布后，他怀疑 OpenAI 会“削弱”旧模型的能力；随后他观察到 ChatGPT 开始出现漏看提示、忽略历史聊天和既有流程、甚至给出错误命令，导致自己在维护本地模型时出错。作者认为旧模型的跨会话检索能力也明显被削弱。
他的结论是：旧模型一旦能满足工作流，就值得继续使用；但被迫迁移到新模型会带来额外学习成本和风险。因此他正在把更多工作负载转向自托管，至少输出更可预测，不会在厂商改动后“悄悄变样”。
More from Models
- Benchmark scores drop from 89% to 19% on new evals — how benchmaxxing breaks leaderboard trust — airesearch12 · 2026-09-11
- ChatGPT tells user their question is too hard and to 'accept dumber answers' — phido3000 · 2026-09-11
- Developer Building a Unified Leaderboard of All Model Benchmark Scores — airesearch12 · 2026-09-11
- Rumor claims Kimi faked performance by serving Claude; DeepSeek new model surprises in evals — realsohamparekh · 2026-09-11
- GPT-5.6 writes well but is instantly forgettable, user complains — BasedRaddka · 2026-09-11
- Opus Refuses Protein Research Codebase Over 'Safety' Concerns, Dev Considers Rolling His Own — josephdviviano · 2026-09-11