GeoChallenge (Text Only): leaderboard
Metric: Exact match in percent on all 908 problems on GeoChallenge-small (908 multiple-choice geometry proof problems sampled to match the full set's difficulty profile; one or more correct options per problem), unified prompt, greedy decoding (temperature 0), 16,384 output tokens, no tools; a problem counts only when the selected option set equals the gold set (no-guess protocol), with the problem text only (no diagram); higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 235B A22B 2507 (Thinking) | 75.55 | #253 (Qwen 3 235B A22B 2507) |
| 2 | O3 Mini (2025-01-31) | 67.84 | #232 |
| 3 | O3 (2025-04-16) | 62.67 | #117 |
| 4 | Gemini 3 Pro (Preview) | 51.21 | #64 |
| 5 | Claude Sonnet 4.5 (Thinking) | 36.23 | #138 (Claude Sonnet 4.5) |
| 6 | Gemini 1.5 Pro | 20.59 | #442 |
| 7 | Claude 3.5 Sonnet | 17.62 | #337 |
| 8 | DeepSeek R1 0528 | 16.41 | #217 |
| 9 | GPT-4o | 13.99 | #333 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=geochallenge-text-only · How It Works · Data refreshed daily, snapshot 2026-10-11.