Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard

A team of five Qwen3.8-27B models, using a method called GVS5H, matched the performance of Fable 5 on LiveCodeBench Hard, a benchmark for programming problems. The team used a multi-agent harness, with each model playing a specific role, and achieved a mean score of 9.8 out of 10 on the first ten problems. The experiment was run on OpenRouter, and the total cost was $35.79 for 50 runs, with a median model time of 40 minutes per run. The authors' own harness and a single Fable 5 model scored 9.6 and 8.0, respectively, on the same problems.