ViSTR-Bench: leaderboard
ViSTR-Bench (2026) tests visual spatial-temporal reasoning from continuous cues in dynamic videos, from vehicle motion and ego-motion to predicting where a basketball, golf or billiard shot ends up and whether a Jenga tower stays up. Accuracy in percent on two-option questions (chance 50), higher is better.
Metric: Accuracy (%; averaged over all 1,340 two-option questions in 15 video subtasks across motion perception, spatial relations, outcome prediction and physical dynamics; chance 50, majority-class 57.9). Source: arxiv.org. Saturation forecast: Around 2029. 39 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Thinking) | 62 | #123 (GPT-5.4) |
| 2 | GPT-5.4 (Non-reasoning) | 56.1 | #123 (GPT-5.4) |
| 3 | Claude Opus 4.6 | 55.4 | #102 |
| 4 | Qwen 3.5 27B | 55 | #359 |
| 5 | Claude Opus 4.6 (Thinking) | 54.7 | #102 (Claude Opus 4.6) |
| 6 | Qwen 3.5 397B A17B | 54.4 | #246 |
| 7 | Gemini 3.1 Pro (Preview) | 53.6 | #94 |
| 8 | Qwen 3.5 122B A10B | 53.4 | #331 |
| 9 | Claude Sonnet 4.6 (Thinking) | 53.4 | #161 (Claude Sonnet 4.6) |
| 10 | Gemini 3.1 Flash Lite (Preview) | 52.6 | #418 |
| 11 | MiMo-V2.5 (Non-reasoning) | 52.4 | #354 (MiMo-V2.5) |
| 12 | Intern-S1 | 51.7 | #617 |
| 13 | Claude Sonnet 4.6 | 51 | #161 |
| 14 | Gemini 3 Flash (Preview) | 50.7 | #144 |
| 15 | GLM-4.6V | 49.9 | #701 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vistr-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.