NerfBench 上线：追踪模型发布后是否被降智

Matthew Miller @matthewmillerai原推发布于 2026.09.26

在 X 上打开原推

I'm tired of frontier models getting nerfed after launch. Every release it's the same cycle. Day one the model is incredible. A few weeks later it feels lazier and dumber, everyone swears it got quantized, and the labs say nothing changed. Nobody has receipts. So we built NerfBench. BridgeBench retests models against their own day one results. Same tasks, same prompts. If a model gets worse, you'll see exactly how much. Claude Opus 5.5 and GPT 6 Astra are first up. Right now Opus 5.5 is the best model I've used. NerfBench is how we find out if it stays that way.

这条是引用着另一条写的：x.com/bridgemindai/status/2103825430189109688