DeepSeek V4.1 Flash now leads agentic coding benchmarks. It scored 77.3 against Anthropic’s 66.1 in an October 2026 LiveBench snapshot. Bloomberg Intelligence reports the US-China AI gap has narrowed to 3%. The model uses a Mixture of Experts architecture with 748 billion parameters. It is priced at $0.30 per million input tokens and $1.20 per million output tokens. China’s National Intelligence Law requires disclosure of API data to authorities. DeepSeek previously exposed unsecured databases containing user chat histories. US federal agencies have banned DeepSeek on government devices. The model is open-weight, allowing self-hosting to avoid API risks. Only three of the top fifteen LiveBench models are Chinese. DeepSeek’s API ecosystem and enterprise support lag behind US leaders. China’s AI sector may remain unprofitable until 2030. US chip export controls have not prevented China’s AI progress.
What changed
US models led on agentic coding benchmarks by roughly 9% in May 2026.
What it unlocks
Running a leading agentic coding model on lower-cost, domestic Chinese hardware.
- 3% US-China AI gap
- 77.3 vs 66.1 agentic coding scores
- $0.30 per million input tokens, $1.20 per million output tokens
- 748B total parameters, 552B core, 196B Engram
What you need to act on it
- self-hosting to avoid API data disclosure risks
- accepting legal exposure under China’s National Intelligence Law for API use
Sources