Stanford AI Index 2026: The Numbers That Matter for Product and Engineering Leaders
Key takeaways
Organisational AI adoption reached 88%; generative AI hit about 53% population adoption within three years — faster than the PC or the internet.
AI agents rose from 12% to about 66% task success on OSWorld, yet still fail roughly one task in three; deployment in business functions remains in single digits.
US private AI investment reached $285.9 billion in 2025, and 1,953 AI companies received new funding in the US alone.
Documented AI incidents rose to 362 from 233, while responsible-AI reporting stays patchy.
Stanford's AI Index Report 2026, published in April by the Institute for Human-Centered AI, is the most widely cited independent dataset on artificial intelligence. Its co-chairs summarise the year in one line: a field that is scaling faster than the systems around it can adapt. Below are the findings with the most direct consequences for companies that build software.
Capability: still accelerating, still jagged
Industry produced more than 90% of notable frontier models in 2025. Several now meet or exceed human baselines on PhD-level science questions, multimodal reasoning and competition mathematics.
On SWE-bench Verified, a benchmark of real software engineering tasks, performance went from 60% to close to the human baseline in a single year.
AI agents improved from 12% to roughly 66% task success on OSWorld, which measures real computer tasks across operating systems — but still fail about one attempt in three.
The frontier remains uneven: a model that earned a gold medal at the International Mathematical Olympiad sits next to a top model that reads an analogue clock correctly only about half the time. Robots succeed in only 12% of household tasks.
The practical reading: capability on well-specified tasks is high and rising, but reliability on open-ended, multi-step work is not yet at a level where unsupervised automation is safe.
Models are converging
Leading models are now hard to tell apart, and open-weight models are more competitive than ever. The performance gap between the best US and Chinese models has effectively closed — as of March 2026 the top model led by just 2.7%. At the same time, benchmarks are saturating, frontier labs disclose less, and independent testing does not always confirm developer claims.
For product teams this lowers the value of betting on a single provider and raises the value of evaluation: test models on your own tasks and keep the integration replaceable.
Adoption is fast and broad
Organisational adoption reached 88%.
Generative AI reached about 53% population adoption within three years, faster than the personal computer or the internet. Adoption correlates with income but has outliers: Singapore at 61% and the United Arab Emirates at 54%, while the US ranks 24th at 28.3%.
The estimated annual value of generative AI tools to US consumers reached $172 billion by early 2026.
Yet adoption is not the same as deployment. The report notes that AI agent deployment remains in single digits across nearly all business functions.
Investment concentrates in the US
US private AI investment reached $285.9 billion in 2025, more than 23 times China's $12.4 billion (the report cautions that Chinese state guidance funds are not fully captured). The US also led in new company formation, with 1,953 newly funded AI companies — more than ten times the next country. However, the number of AI researchers and developers moving to the US has fallen 89% since 2017, including an 80% drop in the last year.
Productivity and work
Studies collected in the report show productivity gains of 14% to 26% in customer support and software development, with weaker or even negative effects in tasks requiring more judgement. In software development, where gains are clearest, US developers aged 22 to 25 saw employment fall nearly 20% from 2024, while employment of older developers continued to grow.
Risk is rising faster than reporting
Documented AI incidents rose to 362, up from 233 in 2024. Almost all frontier developers report capability benchmarks, but reporting on responsible-AI benchmarks remains inconsistent. Research cited in the report also finds trade-offs: improving one responsible-AI dimension, such as safety, can reduce another, such as accuracy.
Public opinion is split. 73% of AI experts expect AI to have a positive effect on how people do their jobs, compared with 23% of the public. Globally, the EU is trusted more than the US or China to regulate AI effectively.
Infrastructure and environment
The US hosts 5,427 data centres, more than ten times any other country, and nearly every leading AI chip is fabricated by a single company, TSMC. AI data centre power capacity reached 29.6 GW — comparable to New York State at peak demand.
What to take from it
For companies building with AI in 2027, the index supports a few working assumptions:
Model choice matters less than evaluation and integration. Leading models are close; measuring them on your own tasks is what differentiates.
Agents need supervision. A one-in-three failure rate on structured tasks means human approval for consequential actions is still a design requirement.
Adoption is not value. High usage coexists with single-digit deployment in business functions; the gap is engineering and process work.
Incidents and regulation will grow together. Logging, access control and documentation are cheaper to build in than to retrofit.
Sources
Stanford Institute for Human-Centered AI, AI Index Report 2026, April 2026 — top takeaways and chapters on technical performance, economy and responsible AI.