DigiNews

Summary

CursorBench 3.1 benchmarks agents on multi-file tasks drawn from real Cursor sessions, comparing models like Fable 5 Max, GPT-5.5, Opus, Sonnet, and GLM against average cost per task. It provides a cost-per-task and token usage breakdown, highlighting price-performance tradeoffs and changes to scoring criteria. The article also explains the methodology based on per-million-token pricing.