An open model beats the frontier on ML engineering under cost limits

Weco AI co-founder Zhengyao Jiang reported benchmarking seven frontier models on three categories of autoresearch task, ML engineering, harness/prompt engineering and algorithmic discovery, with Fable-5 winning overall even under a cost constraint and the open Kimi-K2.7-Code surpassing every frontier model on ML engineering (post). The winner, Claude Fable 5, had just been pulled from public access by US export controls, so the benchmark crowned a model nobody outside Anthropic could run while an open Chinese model held the track closest to the recursive loop. It reads as a measurement of who can do the ML engineering that recursive self-improvement depends on: it follows PostTrainBench’s test of whether agents can post-train themselves and Anthropic’s 52x training-code speedup, and precedes Weco’s own first evidence of consistent recursive self-improvement a month later.