Every spring, Stanford University's Institute for Human-Centered Artificial Intelligence (HAI) publishes a report that sums up the previous year of AI in numbers. The AI Index 2026, released in April 2026, covers 2025. I turned the findings that caught my attention most in the several-hundred-page report into charts, one by one.
1. Capabilities keep accelerating
On SWE-bench Verified, which measures the ability to solve real GitHub issues, success went from 60% to close to 100% in a year. On OSWorld, which measures using a computer like a human, agents' task success rose from about 12% to 66%. On PhD-level science questions, competition maths and multimodal reasoning, models now reach or exceed human level.
This speed has a consequence: benchmarks go stale. Once a test saturates within a year, it stops showing the difference between models. It also explains why the METR experiments I described in the previous post moved towards real tasks instead of lab tests.
2. But the frontier is jagged
This is my favourite finding in the report. In the same year, Google's Gemini Deep Think reached gold-medal level at the International Mathematical Olympiad. Meanwhile, the best model can read an analogue clock correctly only 50.1% of the time. In other words, a coin toss.
Solve olympiad-level maths problems, answer PhD-level science questions, fix real software bugs.
Reading an analogue clock (50.1%). And on structured agent tests, one in three attempts still fails.
For anyone building products, the practical meaning is this: a model being good at a hard task doesn't guarantee it will be good at a neighbouring task that looks easy. AI features need to be tested one by one, with real data.
3. Where is the money going?
The investment picture may look one-sided, but the report carries a warning for the US: the migration of AI researchers and developers into the country has fallen 89% since 2017, with a drop of 80% in the last year alone. The money is there; but the arrival of the people who turn money into products is slowing.
4. The gap between the US and China has closed
In model performance, the gap between the two countries has effectively closed. Since the start of 2025 the lead has changed hands several times; as of March 2026, the best US model (Anthropic's) is ahead by only 2.7%. The rest of the picture shows the two countries are strong in different areas:
- Number of top-tier models
- High-impact patents
- Private investment and new companies
- Data centre capacity
- Academic publications and citations
- Total number of patents
- Industrial robot installations
- Publicly owned AI supercomputers
5. The race for AI sovereignty
Countries are moving to build their own compute infrastructure. Here is the picture for AI supercomputers that are publicly owned or built through public-private partnerships:
Although model production is still concentrated in the US and China, the report says the rest of the world's open-source contributions on GitHub are approaching US levels. That's an important signal for countries like Turkey: building a frontier model takes billions; but contributing to the open-source ecosystem and building products on top of these models is open to everyone.
6. Adoption at record speed
Generative AI reached 53% of the population in three years. 88% of organisations use AI in some form. An interesting detail: the US, the clear leader in investment, ranks 24th in population adoption at 28.3%. The report estimates the value US consumers get from AI at 172 billion dollars a year as of early 2026.
We see the "high adoption, low transformation" finding from the MIT report I covered in the previous post here too: 88% enterprise usage doesn't mean 88% enterprise gains.
7. Education lags behind usage
As someone who taught AI fundamentals to students at Deneyap, this picture feels familiar. Students reach the tool long before the rules do. The real question isn't whether AI will be used, but how. And in most places there is still no framework for teaching that.
8. Responsible AI is falling behind
The report stresses that safety-testing disclosure by model developers is still patchy, and that improving one dimension of responsible AI sometimes worsens another. In other words, there is no common language for measuring safety yet.
9. Experts and the public are not looking in the same direction
This 50-point gap may be the number that people building AI products should care about most. The trust gap between those who build the technology and those who use it doesn't show up in any benchmark, but it decides the adoption of every product.
10. My summary
When I finished the report, three sentences stayed with me:
- The capability problem is largely being solved; the application problem isn't. Models get much better every year, but organisations' ability to turn that capability into value isn't growing at the same speed.
- The frontier is jagged. Where a model shines and where it stumbles can sit side by side. No feature should be trusted without testing it on real user data.
- Trust is the new bottleneck. With incidents rising and the public looking on with suspicion, building safe and transparent products is becoming more than a requirement; it's turning into a competitive advantage.
Sources
- Stanford HAI, The 2026 AI Index Report, April 2026.
- Stark Insider, Stanford 2026 AI Index Report, April 2026.
- Burges Salmon, AI in 2026: what does Stanford's AI Index tell us?, 2026.