AI Index Report 2025 — SDGCounting Document Briefs

Publisher: AI Index Steering Committee, Stanford Institute for Human-Centered Artificial Intelligence (HAI)
Edition: Eighth edition, April 2025; the annual benchmark of the field, published since 2017
Focus: A data-driven read on AI research, technical performance, responsible AI, economy, science and medicine, policy, education, and public opinion
Related: International AI Safety Report 2026, the risk-focused companion to this state-of-the-field census

The AI Index is the most cited annual audit of where artificial intelligence actually stands, assembled from globally sourced data rather than vendor claims. Its 2025 edition documents a field where capability is surging and cost is collapsing at the same time, while the tools meant to measure responsible AI lag far behind. For anyone tracking who benefits from AI and who is left out, the report is a ledger of a widening divide.

The headline

Performance on the hardest new benchmarks jumped sharply in a single year, even as the cost of running a capable model fell more than 280-fold. The United States still produces the most top models, but Chinese systems have nearly closed the quality gap, and private investment continues to concentrate in a handful of countries. The report frames 2024 as the year AI moved from the lab into daily economic and public life.

Yolanda Gil and Raymond Perrault, Co-directors, AI Index: “As high-performing, low-cost, and openly available models proliferate, AI’s accessibility and impact are set to expand even further.”

The numbers

Benchmarks leapt. In one year, scores rose 18.8, 48.9, and 67.3 percentage points on the MMMU, GPQA, and SWE-bench tests; on SWE-bench, systems went from solving 4.4% of coding problems to 71.7%.
Cost collapsed. Querying a model at GPT-3.5 level fell from $20.00 to $0.07 per million tokens between November 2022 and October 2024, a more than 280-fold reduction. Hardware costs dropped about 30% a year.
US leads on model count. US-based institutions produced 40 notable AI models in 2024, to China’s 15 and Europe’s three, but the US–China performance gap on major benchmarks narrowed to near parity.
Investment concentrated. 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; global corporate AI investment hit $252.3 billion.
Incidents climbed. Reported AI-related incidents rose to 233 in 2024, a record high and a 56.4% increase over 2023.
Health uptake accelerated. The FDA had approved 223 AI-enabled medical devices by 2023, up from six in 2015, and a leading model set a state-of-the-art 96.0% on the MedQA clinical benchmark.
Optimism split by region. Majorities in China (83%), Indonesia (80%), and Thailand (77%) call AI more beneficial than harmful, against 39% in the United States and 36% in the Netherlands.

The counting angle

Two of the report’s findings sit squarely in SDGCounting territory. First, the global AI divide is measurable and growing: investment, model production, and infrastructure cluster in a few countries, while access in much of Africa and Latin America is constrained by gaps as basic as electricity. Second, the instruments for responsible AI are thin. AI incidents are rising sharply, yet standardized responsible-AI evaluations remain rare among major developers, and the public data commons that trains these systems is shrinking, with restricted tokens in one common web dataset jumping from roughly 5–7% to 20–33% in a year. What is easy to count, compute and capital, is racing ahead of what is hard to count, harm and accountability.

Our read: The AI Index is strongest where the data is cleanest, and that is exactly the problem it surfaces. Capability and cost are now tracked to the decimal, while safety, bias, and who is excluded are still measured by patchy, voluntary benchmarks. The gap between those two ledgers is where the governance fight will be won or lost.

Watch & read

AI Index Report 2025, the full report (Stanford HAI).
International AI Safety Report 2026, the risk-focused companion on frontier AI and its measurement problem.

Figures are as reported by the AI Index 2025, drawn from its Top Takeaways and chapter highlights; most metrics describe calendar year 2024 unless the report states otherwise.