US Lead in AI Benchmarks Narrows to 3%

Bloomberg Intelligence says the US AI performance lead over China has narrowed sharply to a record low.[1] Its senior analyst Robert Lea puts top Chinese models 3% behind US rivals on benchmark scores.[1] The gap was about 9% in May and 15% earlier in the year.[1] Lea attributes the improvement to DeepSeek’s September release of V4.1 Flash.[1] The figures describe benchmark performance; they do not establish that Chinese models have caught up across all uses or measures of AI capability.[1]

Bloomberg Intelligence says V4.1 Flash ranked sixth globally on LiveBench in September, making it the highest-ranked Chinese model on the benchmark since DeepSeek’s R1 reasoning model drew attention in 2025.[2][3] The model scored 81.1, compared with 83.4 for an Anthropic model.[3] Lea describes its performance as comparable to leading systems from Anthropic and OpenAI.[2] Still, only three of LiveBench’s 15 top-ranked models were Chinese, and Lea cautions that benchmark rankings can change.[2]

DeepSeek says V4.1 Flash is live on its API with native multimodal support, and describes it as the smallest model in its new architecture family, with native visual understanding.[4] The company says it uses a 552-billion-parameter mixture-of-experts architecture, with 8 billion active parameters for input and 16 billion for output.[4] DeepSeek also says it uses one-quarter as much HBM and one-eighth as much SSD storage for its KV cache.[4] Its stated off-peak API rates are half of peak rates.[4]

Lea says better performance could bring further market-share gains for Chinese competitors, but cautions that strong leaderboard scores may be difficult to monetize.[1][2] He says Chinese models face growing US regulatory scrutiny and potential bans following allegations of model distillation.[2] Lea says sustainable profits would require less competitive pressure, industry consolidation and more rational pricing, and says China’s AI industry could remain unprofitable until 2030.[2]