Bloomberg Intelligence: China’s leading AI models trail U.S. by about 3% on benchmarks

Bloomberg Intelligence says the gap between leading U.S. and Chinese AI models has shrunk fast, with Chinese systems now only about 3% behind on benchmark scores after DeepSeek’s V4.1 Flash came out in September. That’s down from roughly 9% in May and 15% earlier this year, a steep change in just a few months.

The shift matters because it suggests Chinese labs are getting better at squeezing more out of their models, especially on hardware built for the domestic market. Robert Lea, Bloomberg Intelligence’s senior analyst, says that progress should translate into more market share for Chinese contenders. He also argues it raises fresh doubts about whether U.S. export controls on chips such as Nvidia’s can really slow China’s AI push.

DeepSeek’s V4.1 Flash ranked sixth globally on LiveBench in September, the highest position for a Chinese model since the startup’s reasoning model R1 broke through in 2025. LiveBench scores models on how they answer and analyze questions, puzzles and tasks. DeepSeek’s last recorded score was 81.1, below Anthropic’s top score of 83.4, but Lea still described it as delivering comparable performance to leading systems from Anthropic and OpenAI.

But the benchmark climb doesn’t make the business case easier. Lea notes that only three of the top 15 LiveBench models were Chinese, and that rankings can move around. He also says Chinese AI may stay unprofitable until 2030, squeezed by low-margin token supply, a brutal price war and a crowded domestic market with more than 1,100 large language models. ByteDance’s Doubao is the monetisation leader, while DeepSeek and Tencent’s chatbots remain free.