SemiAnalysis: Claude Fable 5.2 is token-inefficient on some GDPx tasks

SemiAnalysis says Claude Fable 5.2 is very token-inefficient for certain tasks because of its greater persistence and its tendency to spawn subagents.

On GDPx materials science tasks, it says Fable 5.2 costs 89% more than GPT-6 Astra and uses 4.3 times as many output tokens for effectively the same score, and that the cost gap against Gemini 4 Carbon is 7.5 times. In a second part of the thread, it says Fable 5.2 uses 2.3 times as many output tokens as Claude Opus 5.5 on GDPx finance tasks for a marginally worse score.

These are SemiAnalysis's own figures.