Claude Opus 4.6 seems to have really become less intelligent lately...

Last week, in the BridgeBench hallucination benchmark test, it was still firmly ranked second with an accuracy of 83.3%
As a result, on April 12th, after retesting, it dropped directly to 10th place, with an accuracy of only 68.3%, and hallucination rate skyrocketed by 98%
A comparison chart shows a very obvious gap between the before and after
Many people have recently felt that it has become noticeably dumber when writing code or doing reasoning, forgetting instructions quickly and increasing nonsense