The AI Index Report 2026 is the ninth annual edition of the AI Index, an independently sourced survey of AI progress produced by Stanford HAI (the Stanford Institute for Human-Centered Artificial Intelligence) (Source: aiindex.stanford.edu). The 2026 edition runs 423 pages across 9 chapters, covering research and development, technical performance, responsible AI, the economy, science, medicine, education, policy and governance, and public opinion. The report frames its 15 top takeaways around the claim that AI capability is accelerating rather than plateauing while responsible-AI practice, labor markets, and public understanding lag behind. Data partners named in the report include Epoch AI, GitHub, Lightcast, LinkedIn, McKinsey & Company, Quid and Zeki.
Top takeaways
The report's fifteen top takeaways are, in summary:
- AI capability "is not plateauing. It is accelerating and reaching more people than ever." Industry produced over 90% of notable frontier models in 2025; several models meet or exceed human baselines on PhD-level science questions, multimodal reasoning and competition mathematics; SWE-bench Verified performance rose from 60% to near 100% of the human baseline in one year; organizational adoption reached 88%; and 4 in 5 university students use generative AI.
- The US–China model performance gap "has effectively closed," with the lead changing hands multiple times since early 2025.
- The United States hosts the most AI data centers, whose chips are mostly fabricated by one Taiwanese foundry.
- Models win gold at the International Mathematical Olympiad but cannot reliably tell time, an example of what researchers call the jagged frontier.
- Robots still fail at most household tasks while excelling in controlled environments.
- Responsible AI is not keeping pace with capability, with safety benchmarks lagging and incidents rising.
- The United States leads in AI investment, but its ability to attract global talent is declining.
- Adoption is spreading at historic speed, and consumers derive substantial value from tools often accessed for free.
- Productivity gains appear in many of the same fields where entry-level employment is starting to decline.
- AI's environmental footprint is expanding alongside its capabilities.
- AI models for science can outperform human scientists, though bigger models do not always perform better.
- AI is transforming clinical care, but rigorous evidence remains limited.
- Formal education is lagging behind AI, while people learn AI skills at every stage of life.
- AI sovereignty is becoming a defining feature of national policy, while open-source development redistributes participation.
- Experts and the public hold very different views of AI's future, and global trust in institutions to manage AI is fragmented.
Research and development (Chapter 1)
The report states that AI capability is accelerating rather than plateauing, with industry producing more than 90% of notable frontier models in 2025; using Epoch AI's curated dataset, it puts industry's share of notable models at 91.6%. Fewer notable models were released in 2025 than in 2024. The United States produced 50 notable models, China 30 and South Korea 5. The top organizational contributors in 2025 were OpenAI (19), Google (12) and Alibaba (11); since 2014 Google has produced the most notable models, followed by Meta and OpenAI. The report describes the most capable models as now the least transparent: training code, parameter counts, dataset sizes and training duration are no longer disclosed for several of the most resource-intensive systems, including those from OpenAI, Anthropic and Google. Reported parameter counts have stayed near 1 trillion for three years as frontier-lab reporting stopped, while independently estimated training compute has continued to rise. Synthetic data is still not replacing real data in pre-training; OLMo 3.1 Think 32B, with nearly 90 times fewer parameters than Grok 4, achieves comparable results on several benchmarks through pruning, deduplication and curation.
Global AI compute capacity grew about 3.3x per year since 2022 to approximately 17.1 million H100-equivalents, with Nvidia accounting for over 60% and Google and Amazon supplying much of the remainder and Huawei a small but growing share. The United States hosts the most AI data centers at 5,427, more than ten times any other country, and consumes more energy than any other country. Almost every leading AI chip is fabricated by a single Taiwanese foundry (TSMC), which the report characterizes as making the global hardware supply chain dependent on one foundry, though a TSMC US expansion began operations in 2025.
China leads in publication volume, citations, patent grants and output, and industrial robot installations, while the United States produces more top-tier models and higher-impact patents; South Korea leads in AI patents per capita, and China's count among the top 100 most-cited AI papers rose from 33 in 2021 to 41 in 2024. Open-source development continued to scale, with 5.6 million AI projects on GitHub, Hugging Face uploads tripling since 2023, and US-based projects holding 30 million cumulative GitHub stars among projects above 10 stars. The number of AI researchers and developers moving to the United States has dropped 89% since 2017, and 80% in the last year alone. Switzerland and Singapore lead in AI researchers per capita; no country approaches gender parity, with relatively higher female representation in Saudi Arabia (32.3%), Australia (30.1%) and Canada (29.6%).
Environment
The report documents an expanding environmental footprint for AI. It cites Grok 4 training emissions of 72,816 tons CO₂e and AI data center power capacity of 29.6 GW, which it compares to the peak demand of New York state. It states that annual GPT-4o inference water use may exceed the drinking water needs of 12 million people. Citing Jegham et al. (2025), it reports per-query inference energy for medium-length prompts (about 1,000 input and 1,000 output tokens) of 23 Wh for DeepSeek V3.2 and V3.2 Exp and 21.9 Wh for GPT-5 (high), against 5–6 Wh for Claude 4 Opus and GPT-5 mini (medium); per-prompt emissions ranged from about 14 g CO₂e for the DeepSeek V3.2 models to 1.6 g and 1.5 g for Claude 4 Opus and Mistral Medium 3.
Technical performance (Chapter 2)
The report states that AI capability is outpacing the benchmarks designed to measure it: frontier models gained 30 percentage points in a single year on Humanity's Last Exam, and evaluations meant to be challenging for years are saturated in months. On the SWE-bench Verified coding benchmark, performance rose from 60% to near 100% of the human baseline within one year. Top model performance is converging: as of March 2026, Arena Elo ratings from human voting (LMArena) placed Anthropic (1,503), xAI (1,495), Google (1,494), OpenAI (1,481), Alibaba (1,449) and DeepSeek (1,424) in the top tier, with four companies within 25 points. The open-weight gap reopened in 2025: the top closed-weight model led the top open-weight model by 3.3% in March 2026, up from 0.5% in August 2024, and six of the top ten Arena models are closed-weight.
The report describes the US-China model performance gap as having effectively closed, with models trading the lead multiple times since early 2025. DeepSeek-R1 briefly matched top US models in February 2025, and the report puts the leading US model (Anthropic) ahead by just 2.7% as of March 2026, with the gap fluctuating in the single digits over the year.
The report raises reliability concerns about benchmarks themselves: a review found invalid-question rates from 2% on MMLU Math to 42% on GSM8K, and separate research suggests Arena standing may partly reflect adaptation to the platform. Google DeepMind's Veo 3, tested across more than 18,000 generated videos, showed abilities such as simulating buoyancy and solving mazes without training on those tasks. In professional domains (tax, mortgage processing, corporate finance and legal reasoning) models score 60–90%, with the top 15 models within about 3 percentage points of each other on each benchmark.
The report describes a "jagged frontier" of uneven capability: AI wins gold at the International Mathematical Olympiad yet reads analog clocks correctly only 50.1% of the time. Gemini Deep Think scored 35 points (gold) at the 2025 IMO working end to end in natural language within the 4.5-hour limit, up from a 28-point silver in 2024. (The chapter body gives the top ClockBench score as 50.6%, for GPT-5.4 High in March 2026, against 90.1% for humans.) AI agents improved from 12% to roughly 66% task success on OSWorld (66.3%, within 6 percentage points of human performance) while still failing about one in three attempts on structured benchmarks. Robots fail at most household tasks, with a 12% success rate, even as they reach 89.4% success on RLBench manipulation in controlled lab simulations. Autonomous vehicles reached mass-scale deployment in 2025: Waymo reached approximately 450,000 weekly trips across five US cities, and Apollo Go completed 11 million fully driverless rides in China, a 175% year-over-year increase.
Responsible AI (Chapter 3)
The report states that responsible AI is not keeping pace with capability. Almost all leading developers report capability benchmarks such as MMLU and SWE-bench, but reporting on responsible-AI benchmarks remains sparse. Documented AI incidents in the AI Incident Database rose to 362 in 2025, up from 233 in 2024. It notes that improving one responsible-AI dimension, such as safety, can degrade another, such as accuracy, a trade-off relevant to evaluation frameworks including FMF cyber thresholds and safety cases; it reports that recent studies found training techniques aimed at one dimension (safety, fairness or privacy) consistently degraded others.
Other chapter findings: on a new benchmark distinguishing knowledge from belief, hallucination rates across 26 top models ranged from 22% to 94%, with GPT-4o's accuracy dropping from 98.2% to 64.4% and DeepSeek R1's from over 90% to 14.4% when a false statement was framed as the user's own belief. The average score on the Foundation Model Transparency Index fell to 40 in 2025 after rising from 37 to 58 between 2023 and 2024. On AILuminate, several frontier models received "Very Good" or "Good" safety ratings under standard use, but safety performance dropped for all tested models under jailbreak attempts. On HELM Arabic, a regionally developed model outscored GPT-5.1 and Gemini 2.5 Flash, and on a Slovenian commonsense test several leading models lost close to half their accuracy in a regional dialect. Among organizations, AI-specific governance roles grew 17% in 2025, the share of businesses with no responsible-AI policies fell from 24% to 11%, and the main obstacles were knowledge gaps (59%), budget constraints (48%) and regulatory uncertainty (41%). GDPR remained the most cited regulatory influence but slipped from 65% to 60%; ISO/IEC 42001 was cited by 36% and the NIST AI RMF by 33%.
Economy and labor (Chapter 4)
Global corporate AI investment more than doubled in 2025. Global private AI investment reached $344.7 billion, up 127.5%, of which generative AI companies took $170.9 billion, nearly half and up more than 200%. The report records US private AI investment of $285.9B in 2025, 23 times China's $12.4B, while noting that private figures likely understate Chinese spending given government guidance funds that deployed an estimated $184 billion into AI firms between 2000 and 2023. It counts 1,953 newly funded US AI companies in 2025, ten times the next country; newly funded AI companies worldwide rose 70.8%, the average private investment event rose 46% to $66.5 million, and 28 events exceeded $1 billion, up from 15. It states that the US ability to attract global talent is declining. Leading frontier companies are reaching revenue scale quickly while compute spending rises, and Google reported more than $150 billion in annual capital expenditure in 2025.
Organizational AI adoption reached 88%, with generative AI used in at least one business function at 70% of organizations, while AI agent deployment remains in single digits across nearly all business functions. Generative AI reached 53% population-level adoption within three years, which the report describes as faster than the personal computer or the internet; adoption correlates with GDP per capita, with Singapore at 61% and the United Arab Emirates at 54%. The United States ranks 24th at 28.3% adoption. The consumer value of generative AI in the United States was estimated at $172B annually by early 2026, up 54% from $112 billion a year earlier, with median value per user tripling between 2025 and 2026.
The report links productivity gains of 14-26% (in customer support and software development) to declining entry-level employment in the same fields; the chapter gives 14–15% in customer support, 26% in software development and 73% in marketing output, with smaller gains in tasks requiring deeper reasoning and possible long-term learning penalties from heavy reliance. In software development, US developers ages 22-25 saw employment fall nearly 20% from 2024, even as headcount for older developers grew. One-third of surveyed organizations expect AI to reduce their workforce in the coming year, almost half expect little change, and anticipated reductions are highest in service operations, supply chain and software engineering. The report presents this co-occurrence of productivity gains and entry-level employment decline in the same fields as a recurring tension in the labor data, bearing on the displacement predictions associated with Dario Amodei tracked under AI Labor Disruption. China accounted for 54% of industrial robots installed globally in 2024, up from 51.1% in 2023.
Science (Chapter 5)
The report states that AI for science can outperform human scientists, though larger models do not always win. AI-related natural-science publications in Web of Science reached approximately 80,150 in 2025, up from 63,547 in 2024. On ChemBench the best models surpass human expert averages across 2,700+ chemistry questions, yet frontier models score below 20% on paper-scale replication in astrophysics (ReplicationBench), and LLM agents answer Earth-observation questions with 33% accuracy (UnivEarth) with code failing 58% of the time. On end-to-end research tasks the best agent reaches 38.8% on PaperArena against a PhD baseline of 83.5%, and frontier models reach about 17% on BixBench bioinformatics analysis. Aardvark Weather ran a full weather-forecasting pipeline end to end for the first time, and FourCastNet 3 produces a 60-day global forecast in under 4 minutes. Astronomy released its first foundation model, AION-1, trained on over 200 million celestial objects. Sakana's AI Scientist-v2 produced the first fully AI-generated paper accepted at a peer-reviewed (ICLR) workshop, and Google's AI Co-Scientist was validated in three biomedical areas. Most AI foundation models for science come from academic, government and cross-sector collaborations, in contrast with industry-dominated general-purpose AI.
Medicine (Chapter 6)
In molecular biology, smaller models outperformed larger ones: the report cites a 111M-parameter protein model (MSAPairformer) that beat previous leaders on ProteinGym, and GPN-Star, a 200-million-parameter genomics model, outperformed a 40-billion-parameter model. Virtual cell models (Arc Institute's Evo 2, STATE and DeepMind's AlphaGenome) emerged as a new frontier. AI-driven protein research publications rose about 71%, from 2,259 in 2024 to 3,855 in 2025. In clinical care, ambient AI scribes reduced note-writing time by up to 83%, with one hospital system reporting a 112% return on investment, but the report cautions that rigorous evidence remains limited, noting that nearly half of clinical AI studies use exam questions rather than real patient data; a review of more than 500 studies found only 5% used real clinical data. The FDA authorized 258 AI medical devices in 2025, mostly through device-modification pathways, with only 2.4% of devices with clinical studies supported by randomized trial data. Microsoft's AI Diagnostic Orchestrator paired with OpenAI's o3 scored 85.5% on complex published case studies against 20% for unaided physicians. AI Overviews appear on 84% to 92% of health-related Google searches. In a randomized trial of 150 diabetes patients using a medical digital twin, 71% achieved healthy blood-sugar levels over one year while reducing medications.
Education (Chapter 7)
The report states that 80% of US students use AI for schoolwork, while only 50% of schools have AI policies, and just 6% of teachers say those policies are clear. Computer-science enrollment fell 11% at US four-year universities between 2024 and 2025, while master's graduates in AI software-related fields rose 17% from 2023 to 2024. New AI PhDs in the US and Canada increased 22% between 2022 and 2024, with new PhDs going to academia rather than industry, reversing a decade-long trend. More than 90% of countries offer computer science in primary or secondary school; China and the United Arab Emirates mandated AI education from the 2025–26 school year. AI engineering skills are growing fastest in the United Arab Emirates, Chile and South Africa.
Policy and governance (Chapter 8)
The report identifies AI sovereignty as a defining feature of national policy, with national AI strategies expanding, especially in developing economies, and covers developing-country AI strategies and global standards convergence. More than half of newly adopted strategies in 2024 came from emerging economies. State-backed AI supercomputing clusters in Europe and Central Asia grew from 3 to 44 between 2018 and 2025, against 2 to 8 in South Asia, Latin America and the Middle East and North Africa. Through 2024 East Asia and the Pacific had adopted 77 data-localization measures, sub-Saharan Africa 71, Europe and Central Asia 66 and North America 3. AI-related witnesses in US congressional hearings rose from 5 in 2017 to 102 in 2025, with industry's share rising from 13% to 37%. European public AI contract commitments reached about $3.7 billion over 2013–2024 (United Kingdom $1.6 billion, Germany $505 million, France $320 million), and US public investment in AI contracts and grants was about $20.4 billion over 2013–2024, against $285.9 billion in US private investment in 2025 alone. It describes open-source development redistributing participation, with contributions from the rest of the world approaching the United States on GitHub and now outpacing Europe.
Public opinion (Chapter 9)
On public opinion, the report records a 50-point expert-public gap on AI's future: 73% of experts expect a positive job impact versus 23% of the public; the gaps are similar for the economy (69% vs. 21%) and medical care (84% vs. 44%). Sixty-four percent of Americans expect AI to lead to fewer jobs over 20 years, against 39% of experts; experts expect generative AI to assist 80% of US work hours by 2030 versus the public's 10%. Globally, per Ipsos's 2025 survey of 30 countries, the share saying AI products offer more benefits than drawbacks rose from 55% to 59% while the share saying they make them nervous rose to 52%; India saw the sharpest rise in nervousness (+14 points). Fifty-eight percent of employees globally use AI at work regularly or semiregularly, exceeding 80% in India, China, Nigeria, the UAE and Saudi Arabia. It states that global trust in AI governance varies widely by country: the United States reported the lowest trust in its own government to regulate AI, at 31%, against a global average of 54%, and across 25 countries in Pew's 2025 survey a median of 53% trusted the EU to regulate AI effectively, against 37% for the United States and 27% for China. Across all 50 US states, concern about too little regulation outweighed concern about too much (41% vs. 27% nationally).
Key quantitative data points
| Metric | Value |
|---|---|
| US private AI investment (2025) | $285.9B |
| China private AI investment (2025) | $12.4B |
| US:China investment ratio | 23:1 |
| Global private AI investment (2025) | $344.7B (+127.5%) |
| Generative AI private investment (2025) | $170.9B |
| Organizational AI adoption | 88% |
| Population-level generative AI adoption | 53% (3 years) |
| US adoption rate | 28.3% (24th globally) |
| Consumer value of gen AI (US, early 2026) | $172B/year |
| Productivity gains (customer support, software dev) | 14-26% |
| Young developer (22-25) employment decline | ~20% from 2024 |
| AI incidents documented (2025) | 362 (up from 233) |
| AI data center power capacity | 29.6 GW |
| Global AI compute capacity | ~17.1M H100-equivalents (3.3x/year since 2022) |
| US data centers | 5,427 |
| Notable models by country (2025) | US 50, China 30, South Korea 5 |
| US–China top-model gap (March 2026) | 2.7% |
| FDA-authorized AI medical devices (2025) | 258 |
| Foundation Model Transparency Index average (2025) | 40 (from 58 in 2024) |
Relation to other topics
Several of the report's findings bear on recurring topics elsewhere. The 2.7% US-China gap as of March 2026 quantifies the qualitative gap described under the Fast-Follow Problem and in China and the US Are Running Different AI Races. The OSWorld jump from 12% to 66% quantifies agentic capability growth discussed under Agentic AI, while the persistent one-in-three failure rate is consistent with a capability-deployment gap. The report's finding of 53% adoption in three years, faster than the PC or internet, bears on the slow-diffusion thesis in AI as Normal Technology, while its "jagged frontier" of uneven capability cuts the other way. The TSMC single-foundry dependency relates to supply-chain questions in Compute Governance. The 14-26% productivity gains are larger-scale figures alongside the 15% finding of Brynjolfsson et al. discussed under Generative AI at Work.
Provenance
Published in 2026 by the Stanford Institute for Human-Centered Artificial Intelligence (HAI). 423 pages, 9 chapters. Ninth annual edition of the AI Index. Figures on this page are taken from the report's top takeaways, chapter highlights and chapter text in Raw Sources/; benchmark and investment figures are fast-decay and refer to the report's data cut-off (generally March 2026). Confidence is medium because the page rests on a single source.
Relationships
- depends-on: Stanford HAI — the publishing institute
- depends-on: Epoch AI — source of the notable-models and compute data
- supports: US-China AI Competition: Different Races, Different Metrics — supplies the 2.7% March 2026 model-performance gap and the 23:1 private-investment ratio
- supports: Fast-Follow Problem — quantifies the gap the concept describes qualitatively
- related: China and the US Are Running Different AI Races — a second reading of the same US–China performance data
- supports: Agentic AI — the OSWorld rise from 12% to roughly 66% task success, and the persistent one-in-three failure rate
- supports: Jagged Frontier — the IMO gold versus 50.1% ClockBench contrast
- contradicts: AI as Normal Technology — the finding of 53% population-level generative-AI adoption within three years cuts against the slow-diffusion thesis, while the "jagged frontier" of uneven capability cuts the other way
- supports: AI Labor Disruption — the co-occurrence of 14–26% productivity gains with a roughly 20% employment decline among US developers aged 22–25
- related: Labor Disruption Timelines: Who Predicts What and Why — the same labor data read against forecast timelines
- related: Generative AI at Work — the 14–26% productivity figures sit alongside the 15% finding of Brynjolfsson et al.
- depends-on: Compute Governance — the single-foundry fabrication dependency and 29.6 GW of AI data-center power capacity
- related: Safety Cases for Frontier AI — the responsible-AI trade-off between safety and accuracy bears on evaluation frameworks
- related: AI Sovereignty, OSWorld, SWE-bench, Humanity's Last Exam, LMArena (Chatbot Arena), ISO/IEC 42001 — AI Management System, NIST AI Risk Management Framework 1.0