Bongard Problems (Classic): leaderboard
The classic human-designed Bongard problems (Foundalis set, ~400 puzzles): models must induce the visual rule separating six left from six right figures, with an LLM judge grading strict correctness. Distinct from the synthetic Bongard in Wonderland set.
Metric: Accuracy (% correct). Source: github.com. Saturation forecast: Around January 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 55.4 | #59 |
| 2 | Gemini 3 Pro (Preview) | 50.5 | #67 |
| 3 | Gemini 3 Flash (Preview) | 38.2 | #86 |
| 4 | Gemini 2.5 Flash | 16.2 | #247 |
| 5 | Mistral Medium 3.1 | 11.5 | #412 |
| 6 | Ministral 3 14B | 9.8 | #658 |
| 7 | Mistral Medium 3 | 9.3 | #503 |
| 8 | ministral-3B-2512 | 8.8 | #921 |
| 9 | Ministral 3 8B | 8.6 | #690 |
| 10 | Mistral Small 3.2 | 8.3 | #608 |
| 11 | Mistral Medium 3.5 | 6.9 | #368 |
| 12 | Magistral Small 1.2 | 6.6 | #582 |
| 13 | Magistral Medium 1.2 | 6.6 | #429 |
| 14 | Mistral Small 4 | 5.6 | #524 |
| 15 | Mistral Large 3 | 5.1 | #409 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=bongard-problems-classic · How It Works · Data refreshed daily, snapshot 2026-10-24.