VideoOdyssey-V - Action Recognition: leaderboard

VideoOdyssey (2026) tests ultra-long video understanding with human-written multiple-choice questions on videos averaging 109 minutes, each annotated with how long a viewer must watch continuously to answer. This board is accuracy in percent on the visual subset questions about recognizing actions; higher is better.

Metric: Multiple-choice accuracy (%) on the 192 action recognition questions (identify the actions of humans or other objects in the video) of VideoOdyssey-V (human-written questions on 100 videos of 60 to 255 minutes that need visual cues and long continuous viewing; 1,618 in all), 128 frames for proprietary models and each open model's maximum frame setting; higher is better. Source: arxiv.org. Saturation forecast: Around March 2027. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)57.8#99
2Seed 2.0 Pro54.7#222
3Kimi K2.550#266
4Qwen 3.5 27B50#375
5Gemini 2.5 Pro49.5#276
6GPT-5.246.9#185
7Qwen 3 VL 235B A22B44.3#485
8InternVL3.5-38B38#1021
9Qwen 3 VL 32B34.9#622
10Video-KTR-7B33.9
11Video-R133.3#1379
12VideoLLaMA3-7B29.2#2050
13LLaVA-NeXT-Video-DPO (34B)26.6#2489
14LLaVA-OneVision-1.5-8B25#1864
15Phi-4 Multimodal Instruct24.5#4154

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

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