Opus 5.5 is probably still nerfed

An analysis of reported reasoning tokens per request found that Anthropic's Opus 5.5 likely remains nerfed after the November 9 launch of Fable 5.2, with a bootstrap test comparing before and after the Fable release yielding p=0.105. The finding follows an earlier blog post theorizing that Opus 5.5's reduced output stemmed from testing of Fable 5.2, a theory the new data cannot confirm at conventional significance levels.

Opus 5.5 is probably still nerfed A few days ago, I wrote about Opus 5.5 being nerfed https://micjm.com/blog/get-ready-for-fewer-opus-5.5-em-dashes/ . One theory was that this was because of testing Fable 5.2, so I wanted to check again, since Fable has now launched https://www.anthropic.com/claude-fable-and-mythos-5-2 . Here’s a chart of Opus 5.5’s reported reasoning tokens per request, relative to alternative models in the same battles. The vertical line is the Fable release date, November 9: From our data alone, we can only say that it’s likely that Opus is still nerfed. The p-val on the bootstrap test of before and after the Fable launch is p=0.105.