DeepSeek's V4.1 Flash Overtakes Claude Fable 5.1 Max on Agentic Coding Benchmark as US-China AI Gap Shrinks to Record 3%
Summary
DeepSeek's V4.1 Flash has overtaken Anthropic's Claude Fable 5.1 Max on LiveBench's agentic coding benchmark, scoring 77.3 versus 66.1, as the US-China AI gap shrinks to a record-low 3% per Bloomberg Intelligence analyst Robert Lea — though the 748-billion-parameter model's China operations carry data-sharing risks under the National Intelligence Law.
Key Points
- DeepSeek's V4.1 Flash now leads on LiveBench's agentic coding sub-benchmark with a score of 77.3 versus 66.1 for Anthropic's Claude Fable 5.1 Max Effort, as the US-China model gap narrows to a record-low 3% per Bloomberg Intelligence analyst Robert Lea.
- V4.1 Flash is a 748-billion-parameter Mixture of Experts model (552 billion core backbone plus 196 billion new Engram memory parameters) priced at $0.30 per million input tokens and $1.20 per million output tokens at peak, with a 1-million-token context window and native image understanding.
- China's National Intelligence Law (2017) legally compels DeepSeek to hand over API data — including code and prompts — to the government on demand, a risk compounded by Wiz researchers' 2025 discovery of an unauthenticated DeepSeek database exposing chat histories and API keys.