PerceptionComp: leaderboard
Metric: Accuracy (%) on the 1,114 five-choice PerceptionComp questions over 279 high-scene-complexity videos (seven domains), each needing evidence from several temporally separated moments; native video input where supported, otherwise 64 (some GPT APIs 50) uniformly sampled frames; proprietary and thinking models prompted with chain of thought, instruction-tuned open models answering directly; higher is better. Source: arxiv.org. Saturation forecast: Around 2028. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 45.96 | #93 |
| 2 | Gemini 3 Pro | 44.43 | #77 |
| 3 | Gemini 2.5 Pro | 44.34 | #145 |
| 4 | Seed 2.0 Pro | 44.34 | #111 |
| 5 | Gemini 3.1 Pro (Preview) | 43.72 | #54 |
| 6 | O3 | 43.54 | #121 |
| 7 | GPT-5.2 | 40.75 | #105 |
| 8 | Qwen 3 VL 235B A22B (Thinking) | 38.2 | #228 (Qwen 3 VL 235B A22B) |
| 9 | Gemini 2.5 Flash | 38.15 | #237 |
| 10 | GLM-4.5V | 36.69 | #339 |
| 11 | GPT-5 | 36.45 | #91 |
| 12 | Qwen 3 VL 30B A3B (Thinking) | 35.68 | #338 (Qwen 3 VL 30B A3B) |
| 13 | Qwen 3 VL 30B A3B Instruct | 34.38 | #365 |
| 14 | GPT-4.1 | 34.02 | #240 |
| 15 | Qwen 3 VL 235B A22B Instruct | 34.02 | #264 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=perceptioncomp · How It Works · Data refreshed daily, snapshot 2026-10-11.