StreamPro-Bench: leaderboard

Proactive streaming video understanding from StreamPro-Bench (Li et al., 2026): 1,285 question-answer pairs on 577 videos across seven tasks in which a model watching a 1 FPS stream must decide when to respond and what to say. Scored with StreamPro-F1 in percent, which credits a response only when it arrives inside the task time window and a Gemini 2.5 Pro judge accepts its content, with precision and recall penalizing extra and missed responses; this board is the weighted average over the three task dimensions that the paper reports, and higher is better.

Metric: StreamPro-F1 (%; a response counts only when a Gemini 2.5 Pro judge credits its content and it arrives inside the task time window, and precision and recall over the response trajectory penalize extra and missed responses; videos streamed at 1 FPS; the weighted average (W-Avg.) the paper reports over the three dimensions, 1,285 QA pairs in seven tasks). Source: arxiv.org. Saturation forecast: Around 2029. 11 models tracked.

Top models

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=streampro-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.