Spatial Intelligence Index
How well AI models understand the physical world, at planetary scale: the mean of each model's best score on sets the top models have not solved.
Ranking
Each model once, at its best score on each set, 0 to 100, equal weights. DVL-Bench and GS-QA2 Full join the index once 10 models have finished them.
| Rank | Model | Developer | Index | BigEarthNet.txt | GeoChrono | MultiGlobeQA |
|---|---|---|---|---|---|---|
| 1 | Claude Sonnet 5.5 | Anthropic | 70.9 | 57 | 76 | 80 |
| 2 | Claude Opus 5.5 | Anthropic | 70.7 | 60 | 74 | 79 |
| 3 | Claude Haiku 5.5 | Anthropic | 68.0 | 54 | 73 | 77 |
| 4 | Claude Opus 5 | Anthropic | 67.7 | 53 | 71 | 79 |
| 5 | Qwen 3.8 27B | Alibaba (Qwen) | 63.7 | 48 | 69 | 74 |
| 6 | Claude Opus 4.8 | Anthropic | 60.9 | 48 | 59 | 75 |
| 7 | Claude Opus 4.7 | Anthropic | 59.8 | 49 | 54 | 77 |
| 8 | Claude Sonnet 5 | Anthropic | 58.3 | 47 | 55 | 73 |
| 9 | Step 5 Preview | StepFun | 56.9 | 49 | 49 | 73 |
| 10 | Muse Glimmer 30B | Meta | 56.1 | 46 | 50 | 72 |
| 11 | Nemotron 3.5 Super VL 120B | NVIDIA | 55.7 | 47 | 48 | 72 |
| 12 | MiniMax M3.1 Flash | MiniMax | 54.4 | 44 | 47 | 72 |
| 13 | Qwen 3.6 35B-A3B | Alibaba (Qwen) | 50.4 | 45 | 34 | 72 |
| 14 | Claude Haiku 4.5 | Anthropic | 50.3 | 45 | 36 | 70 |
| 15 | Nemotron 3 Nano Omni 30B | NVIDIA | 38.4 | 37 | 31 | 48 |
Provisional
Not run: one set each, so these are not ranked. The index is the mean of the sets each has.
| Rank | Model | Developer | Index | BigEarthNet.txt | GeoChrono | MultiGlobeQA |
|---|---|---|---|---|---|---|
| – | Mistral Large 4 | Mistral AI | 52.6 | 43 | 62 | – |
Models per set
- DVL-Bench: 8 models (not counted yet)
- BigEarthNet.txt: 16 models
- GeoChrono: 16 models
- MultiGlobeQA: 29 models
- GS-QA2 Full: 6 models (not counted yet)
Frequently asked questions
What is the Spatial Intelligence Index?
A single score for how well an AI model understands the physical world: the mean of its best score on sets that still separate models, each from 0 to 100, all weighted equally.
Which sets count?
DVL-Bench, BigEarthNet.txt, GeoChrono, MultiGlobeQA and GS-QA2 Full, once enough models have finished a set (10 or more). A set leaves the index once the top models solve it.
Why is a model missing?
A model is ranked only when it has a finished run on every set that counts. A model one set short is listed as provisional, with its mean over the sets it has, and is never given 0 for the missing one.
