Text models
Claude Fable 5.2 vs. Opus 5.5
Evaluate review quality, completion time, and the cost of usable work.
Design a review test
Give both models the same patch, commands, and acceptance criteria. Count reproducible findings, false positives, and regressions introduced by their proposed fixes.
Price the usable result
Include elapsed time, token charges, and human review. For long tasks, test whether the model can recover after a changed requirement or resume from saved files.
Test prompt
Review this patch. Report only issues reproducible from the input. Include a failing condition and the smallest correction for each.