SpatialWorld - Work: leaderboard
SpatialWorld (2026) runs multimodal models as agents in 3D simulators under vision-only observation, with every environment behind one text action interface and each task checked by a terminal-state verifier. This board covers the 59 work-scenario tasks. Scored by task success rate in percent; higher is better.
Metric: Task success rate (%) on the 59 work tasks; task success rate (%) judged by terminal-state verifiers, vision-only egocentric observation and a unified text action interface, temperature 1.0, 30-turn history, step budget twice the human golden action count plus 10; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 16.9 | #167 |
| 2 | Qwen 3.5 397B A17B | 16.9 | #246 |
| 3 | Gemini 2.5 Pro | 11.9 | #266 |
| 4 | Gemini 3.1 Pro (Preview) | 10.2 | #94 |
| 5 | Gemini 3 Flash | 10.2 | #164 |
| 6 | Kimi K2.5 | 8.5 | #256 |
| 7 | Qwen 2.5 VL 72B Instruct | 8.5 | #846 |
| 8 | Qwen 3 VL 235B A22B Instruct | 8.5 | #522 |
| 9 | Qwen 3 VL 235B A22B (Thinking) | 8.5 | #451 |
| 10 | Seed 2.0 Lite | 6.8 | #289 |
| 11 | GPT-5.4 | 5.1 | #123 |
| 12 | GLM-4.6V | 5.1 | #701 |
| 13 | GLM-4.5V | 3.4 | #744 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=spatialworld-work · How It Works · Data refreshed daily, snapshot 2026-11-13.