Open-source MoE model matches GPT-5.5 on agentic coding benchmarks

GPT-5.5 achieves 16% Pass@8 on MA-ProofBench Level I and 5% on Level II, outperforming most models.