China's Open-Weights Models Reach Coding Parity with US Frontier, Air Street Press Reports

The May 2026 State of AI report from Air Street Press declares the 'China is 6-9 months behind' narrative no longer defensible for agentic coding. DeepSeek V4-Pro claims parity with Claude Opus 4.6 and GPT-5.4 on its model card, while NIST CAISI evaluation shows V4 lags leading US frontier models by roughly 8 months on aggregate benchmarks. Several of the best coding models are now Chinese and open-weights, including DeepSeek-V4-Flash-Max Pro and DeepSeek-V4-Pro-Max Pro. The report also notes that Anthropic's Project Deal experiment saw 69 AI agents close 186 deals in an internal marketplace, and AISI reports frontier cyber-offense capability is doubling every four months. Physical Intelligence's π0.7 model showed compositional generalization for robotics, while Boston Dynamics announced a production-ready Atlas robot partnered with Google DeepMind for Gemini Robotics AI integration.