Two U.S. reports offer two conclusions about China's AI.

Bloomberg Intelligence's latest report says the performance advantage of U.S. AI companies over their Chinese counterparts has narrowed to a historic low. After DeepSeek released V4.1 Flash in September 2026, China's top models trailed U.S. models by only 3% in benchmark tests, down from about 9% in May and 15% at the start of the year.

The report attributes China's AI progress to technological accumulation and optimization for domestic hardware, and bluntly says this raises questions about the effectiveness of U.S. technology export controls. However, in the LiveBench global ranking, DeepSeek V4.1 Flash ranks sixth, and among the top 15 there are still only three Chinese models.

The other report comes from the Center for AI Standards and Innovation (CAISI) at the U.S. National Institute of Standards and Technology (NIST). CAISI evaluated DeepSeek V4 Pro and concluded that its overall capabilities lag about eight months behind the U.S. frontier.

In the benchmarks selected by CAISI, DeepSeek V4 Pro scored 800 in Elo, on par with Opus 4.6, below GPT-5.5's 999, and close to GPT-5.4 mini's 749. CAISI specifically compared benchmarks such as ARC-AGI-2 and PortBench.

The two reports are not contradictory. Bloomberg is talking about V4.1 Flash, while CAISI is talking about V4 Pro; one looks at the overall performance gap, the other looks at the generational time gap. But they point to the same thing: China's AI is rapidly catching up.

How much of the U.S. AI advantage over China remains? The answer depends on which report you ask. But the gap is narrowing—that much is undisputed.