VisReason (Vision-Centric): leaderboard

VisReason (2026) tests whether multimodal LLMs can reason directly from visual evidence in everyday images, with 1,505 questions in 10 categories such as spotting differences, board positions, Sudoku, geolocation and visual rule induction, built so that a text caption of the image is not enough. The score is the average accuracy over the 10 categories in percent, higher is better.

Metric: Average accuracy (%) over the 10 reasoning categories (1,505 questions on everyday vision-centric reasoning: multiple-choice and fill-in-the-blank answers matched by regular expression, open-ended answers judged by GPT-5-mini, bounding boxes matched at IoU above 0.5 with partial credit), zero-shot with chain-of-thought prompt templates, one per answer format; rows not shaded gray in the paper use explicit reasoning (GPT-5 and Gemini 3 Pro at low reasoning effort); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3 Pro (Preview)Low47.5#119
2GPT-5.2Thinking35.5#185
3GPT-5 MiniHigh33.2#364
4GPT-5Low33.2#178
5O4 Mini27.3#368
6Qwen 3 VL 235B A22B (Thinking)26.7#459
7Qwen 3 VL 32B (Thinking)22.1#573
8GPT-5 Nano19#906
9Qwen 3 VL 8B (Thinking)19#910
10Qwen 3 VL 8B Instruct18.8#1029
11Gemini 2.5 Flash16#463
12MiMo-VL-7B-RL15.9#1155
13GPT-4o15.6#751
14Qwen 2.5 VL 32B Instruct15.4#1062
15InternVL3-8B10.9#1782

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=visreason-vision-centric · How It Works · Data refreshed daily, snapshot 2026-11-21.