Stanford’s 2026 AI Index puts numbers on a year of fast progress and uneven results. Companies produced over 90% of the notable frontier models in 2025, and several now reach or pass human baselines on PhD-level science questions, multimodal reasoning and competition math.
The annual report from Stanford HAI covers research, technical performance, responsible AI, the economy, education, policy and public opinion. A handful of its figures stand out.
Benchmarks jump, and some gaps remain
On SWE-bench Verified, a widely used coding test, scores climbed from 60% to close to 100% within twelve months. Agents improved sharply on OSWorld too, a test built around real desktop tasks on different operating systems. Success on those tasks went from 12% to about 66%. Even so, agents still miss roughly one attempt in three on structured benchmarks.
The report also lists odd weak spots. Gemini Deep Think earned an International Mathematical Olympiad gold, yet the strongest model gets analog clock faces right just 50.1% of the time.
The US and China keep trading the lead
The top spot has changed hands between American and Chinese models more than once since early 2025. DeepSeek-R1 drew level with the best US model for a short time in February 2025. By March 2026, Anthropic’s best model was ahead by only 2.7%.
Each side has its own strengths. The US builds more top-tier models and more influential patents. China is ahead on papers, citations, total patents and installed industrial robots. South Korea tops the world in AI patents relative to population.
The US also has 5,427 data centers, over ten times the count anywhere else, and draws the most energy of any nation. Hardware is even more concentrated: a single company, TSMC, makes nearly all leading AI chips at its Taiwan foundry. A TSMC expansion in the US began operating in 2025.
Money and talent move in different directions
Private AI investment in the US hit $285.9 billion in 2025. China’s figure was $12.4 billion, a 23-fold difference. The report warns that China’s real spending is likely higher than the private numbers suggest, because state guidance funds fill part of the gap. The US also counted 1,953 newly funded AI companies, over ten times the next country.
Talent flows tell a less comfortable story. Fewer AI researchers and developers are relocating to the US: the flow is down 89% since 2017, and most of that fall (80%) came in the past year.
Adoption is outrunning the rules
Within three years, 53% of the population was using generative AI, a faster climb than the personal computer or the internet managed. Uptake differs by country and follows GDP per capita. The United Arab Emirates (64%) and Singapore (61%) sit above expectations, while the US ranks 24th at 28.3%. The report estimates generative AI tools were worth $172 billion a year to US consumers by early 2026, with the median value per user tripling between 2025 and 2026.
Schools are behind. Over 80% of US high school and college students now use AI for schoolwork. Yet only half of middle and high schools have a policy, and a mere 6% of teachers find theirs clear.
- Organizational adoption of AI reached 88%.
- Four in five university students use generative AI.
- New AI PhDs in the US and Canada rose 22% from 2022 to 2024, and the extra graduates took jobs in academia, not industry.
Safety reporting lags capability reporting
Almost every top developer publishes capability benchmark results. Responsible AI benchmark results are published far less consistently. The count of documented AI incidents reached 362, compared with 233 in 2024. New research also suggests that gains in one area, safety for instance, can cost something in another, like accuracy.
Experts and the public disagree
Asked about AI’s effect on how people work, 73% of experts see a positive impact, while only 23% of the public agrees. Similar divides show up on the economy and medical care. Confidence in regulators differs by country. Americans showed the lowest trust in their own government to regulate AI, at 31%, and the EU is trusted more than either the US or China.
National AI strategies are spreading, especially among developing economies, and state-backed spending on AI supercomputing is rising alongside them. Model production is still concentrated in the US and China. Open-source work may change that. Contributions from the rest of the world now outpace Europe on GitHub and are approaching the US, which is feeding more linguistically diverse models and benchmarks.
What to check in next year’s edition
Three numbers are worth tracking: whether the US-China gap stays near 2.7%, whether agent success keeps rising from about 66%, and whether more developers publish responsible AI results. For now the 2026 AI Index shows a field where capability climbs quickly while policy, measurement and public trust trail behind.
Source: hai.stanford.edu