PushupBench - R2: leaderboard

Agreement of predicted and annotated repetition counts on PushupBench (2026), 446 long-form exercise videos of 22 to 117 seconds, sampled at 5 frames per second: the coefficient of determination R2, at most 1, where 0 equals always predicting the mean count. Higher is better.

Metric: Coefficient of determination R2 between predicted and true repetition counts (at most 1; 0 matches predicting the mean, negative values mean constant or random output; predictions off by more than 50 excluded), on PushupBench's 446 long-form exercise clips (22-117 s, 11 creators in 10 countries; 27 clips accept several valid counts) sampled at 5 fps up to 112 frames (Claude models capped at 100 frames by the API), 360p frames, greedy decoding, 10 prompt templates; higher is better. Source: arxiv.org. Saturation forecast: Around March 2027. 20 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Flash0.82#175
2Gemini 3 Pro0.7#145
3Gemini 2.5 Pro0.63#276
4Gemini 2.5 Flash0.49#463
5Qwen 3 VL 32B Instruct0.36#606
6Qwen 3 VL 32B (Thinking)0.35#573
7Gemini 2.0 Flash0.25#723
8Qwen 3 VL 235B A22B Instruct0.17#532
9Claude Opus 4.50.09#160
10Qwen 3 VL 30B A3B (Thinking)0.08#719
11GPT-50.01#178
12Claude Sonnet 4.50#261
13Qwen 3 VL 30B A3B Instruct-0.06#895
14Qwen 3 VL 8B Instruct-0.11#1029
15Llama 4 Maverick-0.15#1110

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

Interactive version: theaggregate.ai/benchmark?slug=pushupbench-r2 · How It Works · Data refreshed daily, snapshot 2026-11-21.