Stanford AI Index 2026
Annual report from Stanford HAI. The numbers that matter for builders.
Source: hai.stanford.edu/ai-index/2026-ai-index-report
Performance
- SWE-bench Verified (coding): 60% → near 100% in one year
- OSWorld (agent tasks): 12% → 66% success rate
- IMO gold medal achieved by Gemini Deep Think
- Same model: 50.1% accuracy reading analog clocks
The jagged frontier is real. Gold medal math and can’t tell time.
Money
- US private AI investment: $285.9B (2025)
- China: $12.4B (23x less, though government funds likely understated)
- US newly funded AI companies: 1,953 (10x next country)
- US consumer value from GenAI: $172B/year, per-user value tripled in one year
Adoption
- GenAI reached 53% population adoption in 3 years (faster than PC or internet)
- Organizational adoption: 88%
- US ranked only 24th (28.3%). Singapore 61%, UAE 54%
- Students using AI: 80%+ (US high school + college)
- Schools with AI policies: 50%. Teachers saying policies are clear: 6%
Talent
- AI researchers moving to US: down 89% since 2017, down 80% in last year alone
- New AI PhDs (US/Canada): +22% from 2022-2024
- PhDs went to academia, not industry (reversal of previous trend)
Safety
- Documented AI incidents: 362 (up from 233, +55% YoY)
- Capability benchmarking: almost universal among frontier developers
- Responsible AI benchmarking: spotty
- Trade-off discovered: improving safety can degrade accuracy
Geopolitics
- US-China model performance gap: effectively closed (2.7% current lead for Anthropic)
- TSMC fabricates almost every leading AI chip (single point of failure)
- US has 5,427 data centers (10x any other country)
- Open-source: “rest of world” contributions approaching US levels
- Trust in AI regulation: EU > US > China globally. US trusts own government at 31% (lowest)
What this means for solo builders
The agent performance jump (12% → 66%) validates betting on agent frameworks now. agent-benchmarks tracks these numbers. The SWE-bench jump confirms that coding agents will keep getting better at the infrastructure level.
$285.9B flowing into AI means more tools, cheaper inference, more competition among providers. llm-providers-credits tracks where to get free/cheap access.
The talent drain (-89%) and PhD-to-academia shift means less competition from big labs hiring everyone. More researchers publishing, more open models. Good for small bets strategy.
53% adoption in 3 years means the market exists. The question is which niche. seed-niche-scoring helps pick.