Weco AI autoresearch benchmark

Weco AI benchmarked seven frontier models on three categories of autoresearch task, ML engineering, harness/prompt engineering and algorithmic discovery, and reported the results under a cost constraint (announcement). Its single appearance is the June 2026 leaderboard on which Claude Fable 5 won overall while the open Kimi-K2.7-Code surpassed the frontier on ML engineering. It sits beside PostTrainBench as one of the corpus’s instruments for measuring how well models do the research work that recursive self-improvement runs on.