97%
▲ 10.6 pts vs Nov 2025
MathArena accuracy of GPT-5.2 (high), released December 2025
20252025
What the chart shows
MathArena runs models on newly released maths competitions soon after they take place, to limit the chance that the problems were in their training data. GPT-5.2 at high reasoning effort, released in December 2025, scored 97%, 10.6 percentage points above the 86.4% of Grok 4.1 Fast (Reasoning), released in November.
On answer-based problems like these, the AI Index says leading models now match or beat top human contestants. Writing rigorous proofs is another matter (see the related chart).
Key numbers
| Measure | Value |
|---|---|
| Grok 4.1 Fast (Reasoning), Nov 2025 | 86.36% |
| GPT-5.2 (high), Dec 2025 | 96.97% |
| Change | +10.61 percentage points |
| Share missed by GPT-5.2 (high) | 3.03% |
How to read it
- The AI Index links this figure to MathArena’s HMMT February 2026 competition, where answers are graded automatically.
- The report text gives about 83% for November 2025; the data shows 86.4%, which we use.
- Only two points are published, so this is a one-month change, not a long-term trend.
Source: MathArena Leaderboard, 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.5.5: MathArena: 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