US AI lead over China shrinks sharply as DeepSeek closes in

The US lead in artificial intelligence is getting a lot harder to see. American AI companies’ performance edge over Chinese rivals has dropped sharply, Bloomberg reports, with DeepSeek’s progress doing much of the work.

The numbers behind that shift are stark. According to the Stanford 2026 AI Index, the performance gap between top US and Chinese models had shrunk to 2.7%, or 39 points, by March 2026. In May 2023, that gap stood at more than 1,300 points.

How the gap closed

DeepSeek sits at the center of the story. The Chinese lab launched its R1 reasoning model in January 2025, a release that put Chinese AI squarely on the global radar.

It followed up with the V4 series, which began rolling out in April 2026. The V4 models brought open-weight access and meaningful cost advantages over US alternatives.

Price is where DeepSeek really competes. DeepSeek V4-Pro costs approximately $3.96 per million output tokens, which significantly undercuts its US competitors.

Developers have noticed. As of mid-2026, Chinese models captured between 50-67% of token processing on OpenRouter, a platform that routes developer requests across many different AI models.

Where the US still leads

The scoreboard is not fully tied. US models from companies like OpenAI and Anthropic still hold an edge on specialized benchmarks, particularly complex reasoning and cyber tasks.

A May 2026 evaluation from CAISI found that even the best Chinese models trail by 6-8 months in those critical areas. DeepSeek V4 Pro specifically lagged US frontier models by approximately 8 months, according to CAISI.

Chinese labs beyond DeepSeek are also in the mix. As of mid-2026, Chinese models ranked in the global top 10 alongside US leaders. Moonshot’s Kimi K3 is frequently cited as one of the models pushing Chinese labs closer to the frontier.

The US keeps clear advantages in infrastructure and intellectual property. It accounts for approximately 74% of compute power and leads in high-impact patents.

China, meanwhile, surpasses the US in publication volume and some citation metrics.

Background: a rivalry shaped by chips

The narrowing gap comes amid an intensifying tech rivalry between Washington and Beijing. US export controls on advanced chips were designed to slow China’s progress by limiting access to the hardware needed to train frontier models.

Chinese labs have responded by competing on innovation and efficiency. DeepSeek’s pricing and open-weight releases suggest a strategy built around doing more with less.

What this means

For US AI companies, the pricing pressure may be the most immediate concern. The OpenRouter share data suggests many developers have already chosen cheaper Chinese alternatives.

The key things to watch are the specialized benchmarks. If the 6-8 month lag CAISI identified in complex reasoning and cyber tasks starts to shrink the way the general performance gap did, the remaining US advantages could narrow quickly.

Also worth tracking is whether DeepSeek’s V4 series and models like Kimi K3 keep expanding share on platforms like OpenRouter. Usage data tends to reveal developer preferences before benchmark tables catch up, and right now it points toward cheaper Chinese options gaining real traction.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.