OmniMatBench - QA: leaderboard

OmniMatBench (2026) poses 3,171 expert-curated multimodal materials-science problems, open-ended questions graded against expert key points and calculations graded slot by slot, across 19 subfields. This board is the macro-F1 of the open-ended answers against the expert key points over all QA questions, 0 to 100; higher is better.

Metric: QA macro-F1 (0-100): F1 between the response and the expert key points (precision and recall judged by Gemini 2.5 Flash), macro-averaged over subfields, times 100, over all 1,261 open-ended QA questions in the 19 subfields on OmniMatBench (3,171 expert-curated multimodal materials-science problems in 19 subfields: 1,261 open-ended QA and 1,910 calculation questions); higher is better. Source: arxiv.org. Saturation forecast: Around October 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.744.3#85
2Gemini 3.1 Pro (Preview)40.6#99
3GPT-5.539.3#54
4O337.6#237
5Grok 4.335.4#342
6Qwen 3.6 Plus35#238
7Gemma 4 31B32.1#382
8O4 Mini30.7#368
9Qwen 3 VL 235B A22B27.8#485
10Intern-S1-Pro25.3#585
11GLM-4.6V24.9#710
12GPT-4o21.9#751
13Llama 4 Maverick20.6#1110

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

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