DeepSeek narrows AI gap with US rivals to just 3%, threatening American dominance

Chinese AI models now trail leading US models by just 3%, down sharply from 15% earlier this year, according to Bloomberg Intelligence. DeepSeek V4.1 Flash ranked sixth globally on LiveBench, signalling rapid gains by Chinese AI developers.

The performance gap between leading US and Chinese artificial intelligence models has narrowed to its lowest level yet, according to a new analysis from Bloomberg Intelligence. Chinese AI models now trail US intelligence models by only 3% on benchmark scores, raising new worries about whether US artificial intelligence can keep its lead.

Bloomberg Intelligence senior analyst Robert Lea said in a report on Monday that the gap has dropped from 9% in May and 15% earlier this year. The latest improvement follows the September launch of DeepSeek V4.1 Flash, showing how quickly Chinese AI companies are upgrading their models and hardware.

DeepSeek closes AI performance gap

Bloomberg Intelligence said that DeepSeek V4.1 Flash ranked sixth worldwide on LiveBench in the month, making it the top Chinese model on the benchmark since the DeepSeek R1 reasoning model first drew attention in 2025.

DeepSeeks' latest LiveBench score was 81.1 compared with 83.4 for Anthropic’s model. Robert Lea said that DeepSeek's performance was comparable to leading intelligence systems from companies such as Anthropic and OpenAI.

However, the Chinese AI presence in the models remains small. Only three of the 15 ranked models on LiveBench are Chinese, Bloomberg Intelligence reported.

Chinese AI firms are adapting to hardware restrictions

China's progress is being driven by growing AI expertise and researchers' ability to optimise models for the hardware available within the country, Lea said.

China faces another problem

Even though model performance is improving fast, Chinese artificial intelligence companies face a business problem.

Pragya Singha Roy