EmbodiedBench Manipulation: leaderboard
Low-level tabletop manipulation where the model outputs pick-and-place actions with 3D coordinates, scored across base, commonsense, complex-instruction, visual, and spatial-reasoning splits.
Metric: Avg Score (%). Source: embodiedbench.github.io. Saturation forecast: Estimated already saturated. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 28.9 | #343 |
| 2 | Claude 3.7 Sonnet | 28.5 | #249 |
| 3 | InternVL3-78B | 26.3 | #352 |
| 4 | Claude 3.5 Sonnet | 25.4 | #350 |
| 5 | InternVL3-38B | 22.6 | #405 |
| 6 | Gemini 1.5 Pro | 21.1 | #445 |
| 7 | Gemma 3 12B (IT) | 20.6 | #664 |
| 8 | InternVL2.5-78B | 18 | #431 |
| 9 | Gemma 3 27B (IT) | 17.5 | #523 |
| 10 | Gemini 2.0 Flash | 16.7 | #342 |
| 11 | Qwen 2.5 VL 72B Instruct | 16.2 | #377 |
| 12 | Llama 3.2 90B Vision Instruct | 14.9 | #622 |
| 13 | InternVL3-8B | 11.5 | #620 |
| 14 | Qwen 2.5 VL 7B Instruct | 9.6 | #654 |
| 15 | Gemini 1.5 Flash | 9.6 | #618 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=embodiedbench-manipulation · How It Works · Data refreshed daily, snapshot 2026-10-20.