Across 31 shared benchmarks, Kimi K2 Instruct scores higher on 3 and Qwen3.5 397B A17B on 28. The widest gap is browsecomp, where Qwen3.5 397B A17B scores 69 against 14.1. Both cost $0.60 per million input tokens on tracked API pricing; Kimi K2 Instruct is cheaper on output ($2.50 against $3.60).
Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing. Quoted rates are the price-setter row we currently track for each model — its direct or vendor-official listing where one exists (direct, direct), otherwise the lowest tracked offer.