38.3%
▲ 30.3 pts vs 2024
best accuracy on Humanity's Last Exam in 2025 (Gemini 3 Pro), up from 8% for o1 in 2024
20242025
What the chart shows
Humanity’s Last Exam is a closed-ended test of about 2,700 hard questions across dozens of academic subjects, some with images, built to stay difficult for AI. The best accuracy rose from 8% in 2024 (o1) to 38.3% in 2025 (Gemini 3 Pro), a gain of 30.3 percentage points in a year.
Even the best model still answered 61.7% of the questions wrongly, and the report notes that confident wrong answers remain common.
Key numbers
| Measure | Value |
|---|---|
| 2025: Gemini 3 Pro | 38.3% |
| 2024: o1 | 8% |
| Change, 2024 to 2025 | +30.3 pts |
| Answered wrongly by the best model, 2025 | 61.7% |
How to read it
- Each bar is the best accuracy recorded that year, one model per year, from the HLE leaderboard (lastexam.ai).
- The number of questions (about 2,700) comes from the report; it is not in the data file.
Source: Center for AI Safety et al., 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.4.4: Humanity’s Last Exam (HLE): accuracy. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Discussion