FIVE CHEAP MODELS IN A TRENCHCOAT JUST TIED CLAUDE'S BEST CODER

Researchers at Persis Capital built GVS5H, a harness that runs five fresh copies of Alibaba's open-weight Qwen3.8-27B as a tiny software team — a manager, workers and a verifier — coordinating through shared files on disk instead of talking to each other. No extra training, no fine-tuning, no task-specific demonstrations.

On the 100 most recent hard LiveCodeBench problems, the orchestrated 27B jumped from 66.8 percent on its own to 92.4 percent, edging past Anthropic's Claude Fable 5, which scored 90.4 percent as a single call. The bigger sibling, Qwen3.8-Flash-Next, reached 93.0 percent.

The tie comes with fine print the paper states itself: the manager arm gets up to ten rounds of calls plus a test runner against Fable 5's single call, and the 27B doesn't beat Fable 5 in any statistical sense — the p-value is 0.88. It ties. Which is still the point: organization, not size, did the work.