Stanford 457-Page AI Index Report: Costs Drop, Open-Source Closes Gap
Stanford University released the 457-page 2025 AI Index Report, tracking the latest trends in AI costs, efficiency, model performance, and industry adoption. The report highlights that AI is becoming drastically cheaper and more efficient, industry dominance has solidified, and open-source models are rapidly closing the gap with top-tier closed-source models.
Confirmed
- Cost and Efficiency: The report estimates that AI's price-performance ratio improves by about 30% annually, and energy efficiency improves by 40% annually. Meanwhile, compute demand doubles every five months.
- Open-Source Closing In: As of February 2025, the gap between top open-source and closed-source models on the Chatbot Arena has narrowed to roughly 1.70%.
- Frontier Model Convergence: On the Chatbot Arena, the performance gap between the 1st and 10th ranked models shrank from 11.9% to 5.4% within a year.
- Industry Dominance: In 2024, about 90% of notable AI models originated from the industry, a significant increase from 60% in 2023. Frontier AI is increasingly becoming a competition over products and infrastructure.
Why it matters
These figures indicate that AI technology is marching towards widespread deployment with lower barriers and higher efficiency. The rapid rise of the open-source ecosystem is breaking down the absolute barriers of closed-source models, while surging compute demands and high industry monopoly suggest that the core of AI competition is accelerating towards infrastructure and commercialization capabilities.
2026-07-25 ~ 2026-07-25 · 6 related posts
Primary sources
- [source] Stanford’s 457-page AI Index 2025 report tracks falling costs, efficiency gains and adoption — mdancho84 · 2026-07-25
- Stanford’s 457-page AI report tracks falling costs, better benchmarks and wider adoption — mdancho84 · 2026-07-25
- AI price-performance improves 30% annually, driving mainstream adoption — mdancho84 · 2026-07-25
- [source] Industry dominates 90% of frontier models; training compute doubles every 5 months — mdancho84 · 2026-07-25
- Frontier model performance converges: Gap between #1 and #10 narrows to 5.4% — mdancho84 · 2026-07-25
- [source] Stanford Report: Open-weight models catching up, gap with closed narrows to 1.7% — mdancho84 · 2026-07-25