MA-Bench (Open-Ended): leaderboard
Metric: GPT-4o judge score (0 to 5) averaged over the six open-ended dimensions: descriptive understanding (core action semantics, spatial accuracy, temporal consistency) and reasoning and explanation (body-level label, action-level label, causal reasoning chain), MA-Bench's 1,000 micro-action videos (about 2 seconds each, 52 micro-action categories), 8 sampled frames, zero-shot, final answer without intermediate reasoning; higher is better. Source: arxiv.org. Saturation forecast: Around March 2028. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 0.76 | #237 |
| 2 | GPT-4o | 0.73 | #333 |
| 3 | InternVL2-8B | 0.52 | #826 |
| 4 | Qwen 2.5 VL 7B Instruct | 0.37 | #643 |
| 5 | Qwen 2 VL 7B | 0.31 | #741 |
| 6 | Pixtral-12B | 0.18 | #795 |
| 7 | InternVL3-8B | 0.1 | #606 |
| 8 | Phi-4 Multimodal Instruct | 0.1 | #896 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=ma-bench-open-ended · How It Works · Data refreshed daily, snapshot 2026-10-11.