Four Chinese Labs Ship Open-Weights Coding Models in 12-Day Window at One-Third Frontier Cost

Air Street Press·medium signal

Air Street's May 2026 report documents an unprecedented 12-day window where four Chinese labs released competing open-weights coding models: Z.ai's GLM-5.1, MiniMax M2.7, Moonshot's Kimi K2.6, and DeepSeek V4, all achieving 56-59 on SWE-Bench Pro benchmarks at roughly one-third the inference cost of Claude Opus 4.7. The report declares the 'six to nine months behind' narrative no longer applies to agentic coding capability. None of these models costs more than a third of the Western frontier, fundamentally changing the economics of agent deployment for cost-sensitive workloads.