SemiAnalysis: Claude Fable 5.2 debuts first on its GDPx leaderboard at 16.4%, at 31% higher cost per task than Fable 5.1

SemiAnalysis says Claude Fable 5.2 debuts at the top of its GDPx leaderboard with 16.4%. It says the gains over Claude Opus 5.5 are mostly in chip design (+10.8%) and software engineering (+15.5%), with smaller gains in materials science (+5.3%) and biology (+7.7%), and little change in quant and hardware design.

On the RSI category, which SemiAnalysis uses to track AI-research ability, it says Fable 5.2 scores about the same as Claude Sonnet 5.5. It reads this as consistent with Anthropic's statement that it limited advances in AI research during training.

From reading the model's rollouts, SemiAnalysis says Fable 5.2 is much more thorough in testing its work: it often writes hundreds or thousands of test cases, where Opus 5.5 and Fable 5.1 stopped at about 20. It says this persistence costs more: the average cost per task is 31% higher than Fable 5.1's ($88.15 against $67.16), even though Fable 5.2 is 20% cheaper per million tokens.

SemiAnalysis calls Fable 5.2 a marginal improvement, not comparable to the original Mythos release. These are SemiAnalysis's own figures and opinions.