LiveNerf begins tracking Claude Fable 5.2 performance
The LiveNerf maintainer added Claude Fable 5.2 to a preregistered daily evaluation using a frozen 78-question set from GPQA Diamond, MMLU-Pro, competition math, and AIME. Its first three scores were 80.8%, 88.5%, and 84.6%, with no serving errors, but the panel was selected to be difficult for Claude Opus 5.5 and the two models used different Claude Code versions. The maintainer said an initial ten-day baseline and a rule requiring two consecutive declining ten-day windows mean no long-term stability verdict is possible yet.
The project provides a repeatable, append-only way to track whether model performance changes over time, although its current Fable 5.2 results come from one account and only three days of data.