MeasureBench Real-world: leaderboard

Each model reads the value and unit off 1,272 photographs of measuring instruments (dials, digital displays, linear scales, composites). The score is the percentage of answers whose value falls in the annotated interval and whose unit matches; unlike chart or OCR benchmarks the hard part is locating a pointer or meniscus between tick marks.

Metric: Accuracy (%). Source: raw.githubusercontent.com. Saturation forecast: Around 2029. 13 models tracked. Updated 17 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3 VL 235B A22B Instruct22.6#540
2GPT-5 Mini22#359
3GPT-519.8#177
4Qwen 2.5 VL 7B Instruct14.6#1893
5Qwen 2.5 VL 72B Instruct14.5#849
6Claude Opus 4.1Non-reasoning14.3#279
7InternVL3.5-38BNon-reasoning12.9#1017
8Claude Sonnet 412.6#411
9Qwen 2.5 VL 32B Instruct11.7#1074
10Mistral Medium 3.110.6#865
11InternVL3.5-8BNon-reasoning9.7#1588
12Mistral Small 3.28.5#1645
13Grok 47.5#315

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=measurebench-real-world · How It Works · Data refreshed daily, snapshot 2026-11-17.