OmniCoT: leaderboard
OmniCoT (2026) is a panoramic spatial reasoning benchmark for multimodal models with multi-hop questions over full 360-degree images in three dimensions: See (viewpoint transformation), Locate (inter-object relations) and Move (embodied action simulation). This board is overall accuracy in percent on the synthetic OmniCoT-B panoramas with chain-of-thought prompting; higher is better.
Metric: Overall accuracy (%), question-weighted over the See, Locate and Move dimensions, with chain-of-thought prompting on OmniCoT-B (synthetic 360-degree equirectangular panoramas with three anchor objects given by coordinates and orientation); answers are object identities, visibility, directions or the closest angle option, exact match after normalization. Source: arxiv.org. Saturation forecast: Around 2029. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 39.72 | #167 |
| 2 | Seed 1.8 | 38.9 | #242 |
| 3 | Qwen 2.5 VL 72B Instruct | 35.26 | #846 |
| 4 | Grok 4 | 34.99 | #307 |
| 5 | Qwen 2.5 VL 32B Instruct | 32.39 | #1053 |
| 6 | ChatGPT 4o | 31.58 | #610 |
| 7 | GLM-4.5V | 29.95 | #744 |
| 8 | Gemini 3 Flash | 28.43 | #164 |
| 9 | Qwen 3 VL 8B Instruct | 27.56 | #981 |
| 10 | Qwen 2.5 VL 7B Instruct | 23.53 | #1833 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=omnicot · How It Works · Data refreshed daily, snapshot 2026-11-13.