FlagEval VQA - Visual Puzzles: leaderboard
BAAI FlagEval’s multimodal VQA leaderboard, Visual Puzzles slice: puzzle images whose answer is a pattern rather than a fact.
Metric: Score. Source: flageval.baai.ac.cn. Saturation forecast: Estimated already saturated. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (High) | 44.7 | #98 (GPT-5) |
| 2 | O4 Mini (High) | 40.9 | #180 (O4 Mini) |
| 3 | Gemini 2.5 Pro | 37.9 | #152 |
| 4 | O3 Mini (High) | 37.9 | #272 (O3 Mini) |
| 5 | O4 Mini (Low) | 37.1 | #180 (O4 Mini) |
| 6 | GPT-5 (Low) | 35.6 | #98 (GPT-5) |
| 7 | O3 Mini (Low) | 32.6 | #272 (O3 Mini) |
| 8 | Step-3 | 29.5 | #365 |
| 9 | Gemini 2.5 Flash (Thinking) | 28 | #245 (Gemini 2.5 Flash) |
| 10 | Claude Sonnet 4 (Thinking) | 27.3 | #199 (Claude Sonnet 4) |
| 11 | GPT-4.1 | 26.5 | #248 |
| 12 | Claude Sonnet 4 (Non-reasoning) | 25.8 | #199 (Claude Sonnet 4) |
| 13 | Gemini 2.5 Flash (Non-reasoning) | 23.5 | #245 (Gemini 2.5 Flash) |
| 14 | ERNIE-4.5-VL-424B-A47B | 16.2 | #391 |
| 15 | Mistral Medium 3.1 | 15.9 | #408 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=flageval-vqa-visual-puzzles · How It Works · Data refreshed daily, snapshot 2026-10-20.