SPUR - Numerical Perception: leaderboard

SPUR (2026) tests multimodal models with multiple-choice questions on multi-panel biomedical experimental figures from PubMed Central. Accuracy in percent on numerical perception questions, which ask the model to quantify levels and differences of visual features within a panel; higher is better.

Metric: Accuracy (%) on SPUR the 636 numerical perception multiple-choice questions about multi-panel biomedical experimental figures from PubMed Central papers (questions drafted by GPT-4o, kept only when GPT-4o failed them in at least six of ten text-only attempts, then expert-reviewed): the questions ask the model to quantify absolute levels and differences of visual features within a panel; higher is better. Source: arxiv.org. Saturation forecast: Around August 2027. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3 Pro (Preview)61.26#119
2Claude 3.7 SonnetThinking59.67#470
3O4 MiniHigh59.24#368
4GPT-5.158.73#240
5GLM-4.5V57.7#754
6Gemini 2.5 Pro (Preview 06-05)56.47#213
7Qwen 3 VL 30B A3B (Thinking)53.48#719
8Seed-1.653.3#469
9Llama 4 Maverick51.73#1110
10Ministral-3-8B-Instruct-251251.57#1907
11Ministral-3-14B-Instruct-251250.88#1639
12Grok 4.1 Fast47.33#427
13InternVL3-78B46.3#759
14Qwen 3 VL 30B A3B Instruct44.18#895
15Mistral Small 3.142.74#1706

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

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