China Narrows AI Lead Over U.S. to 39 Arena Points

The Stanford Institute for Human-Centered Artificial Intelligence released its 2026 AI Index this week, reporting that China’s top large-language model trails the U.S. leader by 39 Arena points. The index lists Anthropic’s Claude Opus 4.6 ahead of China’s Dola-Seed 2.0 by 2.7%.

The report compares Arena scores, which measure relative performance of large-language models. In May 2023 the U.S. leader, OpenAI’s GPT-4, led by more than 1,300 Arena points while China’s top model scored below 1,000. By March 2026 the gap had narrowed to 39 Arena points.

Stanford’s index counts 50 notable models developed by U.S. teams and 30 from Chinese teams. The report also records changes in research influence: authors based in China accounted for 20.6% of AI publication citations in 2024, while authors based in the United States accounted for 12.6%.

The index tracks industrial robotics and finds China with more than 295,000 robot installations versus about 34,200 in the United States. It also lists private investment in AI for 2025 at $285.9 billion in the United States and $12.4 billion in China.

On talent flows, the report says the number of researchers and developers relocating to the United States has fallen 89% since 2017. The decline accelerated by 80% in the most recent year measured.

ByteDance’s cloud unit has released an international version of its AI video tool, Seedance 2.0, for business customers in more than 100 countries. The service is not available to American companies. The company delayed the global rollout after AI-generated clips that used recognizable actors and other copyrighted material spread online and prompted legal claims.

ByteDance added safeguards for approved customers: the platform blocks uploads of real human faces, applies filters to prevent use of copyrighted content, offers a library of more than 10,000 computer-generated faces, and requires written permission to use a real person’s likeness. The company said it will tag AI-generated videos using the C2PA standard. The tool runs through a prepaid API on the BytePlus ModelArk platform, accepts text, images, video or audio, and produces MP4 files up to 720p for clips between 4 and 15 seconds. Technical documentation and programming examples are available for approved customers.

ByteDance also responded to a report that it offered nearly RMB 100 million per year to a researcher from another firm. Li Liang, vice president of Douyin Group, wrote on Weibo that the report was “inaccurate” and described hiring terms for the Seed team as a standard package of cash, stock options and equity incentives tied to Doubao. The company says options vest over four years and employees do not face special conditions to receive the full benefits.

The index compiles performance metrics, investment figures, publication and patent counts, and talent flows to map global AI progress. Its 2026 findings record a narrowed performance gap between top U.S. and Chinese language models, higher citation share for Chinese authors in 2024, a larger number of industrial robot installations in China, and a sharp fall in the flow of international AI talent to the United States.