OMIBench: leaderboard

OMIBench (2026) tests vision-language models on Olympiad-level science and mathematics problems whose evidence is spread across several images. This board is the GPTScore in percent, the share of answers an LLM judge finds equivalent to the gold answer, over all problems; higher is better.

Metric: GPTScore (%, binarized LLM-judge equivalence of the final answer with the gold answer, micro-averaged) on the 1,322 problems of OMIBench (Olympiad-level multi-image problems from biology, chemistry, mathematics and physics competitions), zero-shot; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 37 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro (Preview)50.53#123
2Qwen 3 VL 32B (Thinking)49.54#577
3GPT-548.11#185
4GPT-5 Mini47.73#373
5Qwen 3 VL 30B A3B (Thinking)47.2#727
6Qwen 3 VL 235B A22B (Thinking)47.05#462
7O4 Mini45.97#372
8Gemini 2.5 Pro44.1#280
9Qwen 3 VL 8B (Thinking)38.65#924
10Qwen 3 VL 32B Instruct35.78#607
11Gemini 2.5 Flash35.25#466
12Qwen 3 VL 235B A22B Instruct34.39#541
13Qwen 3 VL 4B (Thinking)30.03#1354
14Qwen 3 VL 30B A3B Instruct28.59#903
15Qwen 2.5 VL 72B Instruct27.99#849

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=omibench · How It Works · Data refreshed daily, snapshot 2026-11-22.