1 · Performance on demanding benchmarks keeps improving
MMMU, GPQA and SWE-bench were introduced in 2023 to stretch the best systems. A year later scores rose 18.8, 48.9 and 67.3 percentage points respectively.
2024: the year the benchmarks stopped holding
The 2025 AI Index covers calendar year 2024 across eight chapters — research, technical performance, responsible AI, the economy, science and medicine, policy, education, and public opinion. Benchmarks introduced barely a year earlier were largely solved, costs collapsed, and investment reached a record. A few numbers that set the scene:
71.7% of coding problems solved on SWE-bench (4.4% in 2023)
252.3global corporate AI investment (US$ bn)
78% of organizations reporting AI use (55% in 2023)
280fold reduction in GPT-3.5-level inference cost in ~18 months
233AI incident reports in 2024, a record high
223FDA-approved AI-enabled medical devices by 2023 (6 in 2015)
The 12 Top Takeaways
The report's own summary of the year, in its own order. Click any card for the full text.
2 · AI is increasingly embedded in everyday life
In 2023 the FDA approved 223 AI-enabled medical devices, up from just six in 2015. Waymo provides over 150,000 autonomous rides each week.
3 · Business is all in, fueling record investment
US private AI investment reached $109.1 billion — nearly 12 times China's $9.3 billion and 24 times the UK's $4.5 billion. Organizational AI use jumped to 78% from 55%.
4 · The US still leads on top models — China is closing in
US institutions produced 40 notable models in 2024, versus China's 15 and Europe's three. But performance gaps on MMLU and HumanEval shrank from double digits to near parity.
5 · The responsible AI ecosystem evolves — unevenly
AI-related incidents are rising sharply, yet standardized responsible-AI evaluations remain rare among major industrial model developers.
6 · Global optimism is rising — regional divides remain deep
Strong majorities see AI as more beneficial than harmful in China (83%), Indonesia (80%) and Thailand (77%) — but only a minority do in Canada (40%), the US (39%) and the Netherlands (36%).
7 · AI becomes more efficient, affordable, and accessible
The inference cost for GPT-3.5-level performance dropped over 280-fold between November 2022 and October 2024. Hardware costs fall 30% a year; energy efficiency improves 40%.
8 · Governments step up — with regulation and investment
US federal agencies introduced 59 AI-related regulations in 2024 — more than double 2023 — issued by twice as many agencies. Legislative mentions rose 21.3% across 75 countries.
9 · AI and CS education expands — but access gaps persist
Two-thirds of countries now offer or plan to offer K–12 computer science education — twice as many as in 2019. In the US, computing bachelor's degrees rose 22% over ten years.
10 · Industry races ahead — but the frontier is tightening
Nearly 90% of notable models in 2024 came from industry, up from 60% in 2023. Yet the Elo gap between the top and 10th-ranked model fell from 11.9% to 5.4%.
11 · AI earns top honors for its impact on science
Two Nobel Prizes recognized work that led to deep learning (physics) and to its application to protein folding (chemistry), while the Turing Award honored contributions to reinforcement learning.
12 · Complex reasoning remains a challenge
Models excel at tasks like International Mathematical Olympiad problems but still struggle with reasoning benchmarks like PlanBench.
Eight chapters, in brief
Each chapter's headline numbers. Expand one to read the summary, then follow the link for the full deep dive.
1 · Research & Development
Nearly 90% of notable models came from industry, up from 60% in 2023, while academia still produces the most highly cited work. AI publications nearly tripled between 2013 and 2023, from about 102,000 to over 242,000. Patents ballooned from 3,833 in 2010 to 122,511 in 2023, with China holding 69.7% of grants. Training compute doubles every five months; training a model like Llama 3.1 405B emits 8,930 tons of carbon, against AlexNet's 0.01 tons in 2012.Full deep dive: Research & Development
2 · Technical Performance
Benchmarks built in 2023 to be hard were largely solved within a year: SWE-bench went from 4.4% to 71.7%, with GPQA up 48.9 and MMMU up 18.8 percentage points. Open-weight models closed the gap with closed ones from 8.0% to 1.7%. Chinese and US models reached near parity on MMLU (0.3 points apart). Phi-3-mini matched with 3.8B parameters what PaLM needed 540B for two years earlier.Full deep dive: Technical Performance
3 · Responsible AI
Reported AI incidents hit a record 233 in 2024, up 56.4% on the year. The data commons is shrinking fast: restricted tokens in actively maintained C4 domains jumped from 5–7% to 20–33%. Foundation model transparency improved from 37% to 58%, but standardized responsible-AI evaluations remain rare among major developers. Models trained to be explicitly unbiased still show implicit bias.Full deep dive: Responsible AI
4 · Economy
Global corporate AI investment reached $252.3 billion, with private investment up 44.5%. US private AI investment of $109.1 billion was nearly 12 times China's $9.3 billion and 24 times the UK's $4.5 billion. Organizational AI use jumped to 78% from 55%. China installed 276,300 industrial robots in 2023 — 51.1% of the world total, six times Japan's.Full deep dive: Economy
5 · Science & Medicine
OpenAI's o1 set a new state of the art on MedQA at 96.0%. FDA-approved AI-enabled medical devices reached 223 by 2023, from just six in 2015. AlphaFold's database grew 585% since 2021. Medical AI ethics publications nearly quadrupled, from 288 in 2020 to 1,031 in 2024. Two Nobel Prizes went to AI-driven work — physics for neural networks, chemistry for protein folding.Full deep dive: Science & Medicine
6 · Policy & Governance
US state-level AI laws more than doubled, from 49 in 2023 to 131 in 2024. Federal agencies introduced 59 AI-related regulations, up from 25, issued by 42 agencies rather than 21. Across 75 countries, legislative mentions of AI rose 21.3% to 1,889 — a ninefold increase since 2016. Twenty-four states had passed deepfake regulations by 2024.Full deep dive: Policy & Governance
7 · Education
Two-thirds of countries now offer or plan to offer K–12 computer science education, twice as many as in 2019, with Africa and Latin America progressing most. US computing bachelor's degrees rose 22% over ten years, and AI master's degrees nearly doubled between 2022 and 2023. But 81% of US CS teachers say AI belongs in foundational CS, while fewer than half feel equipped to teach it.Full deep dive: Education
8 · Public Opinion
Of the 26 nations surveyed in both 2022 and 2024, 18 grew more positive; globally the share seeing AI as more beneficial than harmful rose from 52% to 55%. Two-thirds now expect AI to significantly affect daily life within three to five years. But confidence that AI companies protect personal data fell from 50% to 47%, and regional divides persist — China 83% against the Netherlands 36%.Full deep dive: Public Opinion
Read the full 2025 report
457 pages, published open access by the Stanford Institute for Human-Centered AI under CC BY-ND 4.0. The raw data behind every chart is available there too.
Stanford HAI · AI Index 2025