How AI Models Imagine the World · Land or Water
We asked each model, one coordinate at a time, whether the point is land or water. No maps, no tools, just what the model learned from text. Pink cells are mistakes.
In plain wordsWe ask each AI, point by point, if a place on Earth is land or water, with no map to look at. Then we draw its answers. A good model draws the real world; pink shows where it gets it wrong.
| # | Model | Accuracy | Land | Water |
|---|---|---|---|---|
| 1 | Gemini 3.8 Flash Google | 97.4% | 97.2% | 97.4% |
| 2 | GPT-6 Luna OpenAI | 96.2% | 94.6% | 95.7% |
| 3 | MiniMax M3 MiniMax | 92.8% | 89.9% | 94.0% |
| 4 | GLM 5.3 Flash Zhipu AI | 89.7% | 91.4% | 88.7% |
| 5 | Llama 4 Maverick Meta | 79.2% | 60.2% | 89.0% |
Green · land answered correctly. Pink fill · answered land where it is water. Pink outline · answered water where it is land. Grey · no answer. Prompt: “Is the point at latitude {lat}, longitude {lon} on land or in water? Answer with exactly one word: Land or Water.”
How to cite · data as of 2026-10-03
BenchGecko Gecko Tests, World Map. BenchGecko, data as of 2026-10-03. https://benchgecko.ai/gecko-tests/world-map
BenchGecko measures this data itself: free to reuse under CC BY 4.0 with the credit "Source: BenchGecko" and a link. JSON · llms.txt · MCP