Stanford AI report: Model capability accelerating, China has closed the gap with the US
Stanford’s annual AI Index Report says “leading models are now nearly indistinguishable” from each other, and benchmarks used to evaluate them are falling behind their accelerating capabilities.
China has closed the gap with the US in the global competition for AI model superiority, and any talk of an AI performance plateau is just flat-out wrong. Those are two big takeaways from this year’s AI Index Report from Stanford University’s Institute for Human-Centered Artificial Intelligence.
Models are continuing to make impressive leaps in performance, and the pace is accelerating, the 2026 edition of the report finds. But it also carefully examines this rapid acceleration to show that it may not be exactly what it seems.
While last year’s class of AI models have been steadily approaching or surpassing human baseline scores on benchmarks, all of the major frontier models are clustered at the top of the benchmark charts together, all separated by just a few points between them. The authors of the report say “leading models are now nearly indistinguishable from one another.”
That could mean that the ecosystem of benchmarks — the yardsticks used to measure the models’ capabilities — may not be keeping pace with the rapidly evolving models’ skills. For example, the report notes that AI models can win a gold medal in the International Mathematical Olympiad, but are hilariously bad at telling the time by reading a clock face.
And in another worrying sign for its plan to dominate AI, the US has seen a rapid, alarming decline in its ability to attract global AI talent. The Trump administration has added significant new restrictions to the H-1B visa program, which helped lure top AI talent to the US. New requirements such as imposing a $100,000 fee that employers must pay per H-1B hire have resulted in a sharp drop in AI researchers coming to the US.
