GeoBench Photos
Multimodal
GeoBench Photos is the benchmark's 50-location map of unofficial photospheres (real user photos rather than Street View) spanning all countries, scored with GeoGuessr's 0-5000 distance points from a single still image. Atlas imports the mean round score only for runs with no refusals, because the leaderboard averages over answered rounds.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 4314 |
| Claude Fable 5.1 | 4260 |
| Gemini 3.8 Flash | 4251 |
| Gemini 3 Flash Preview | 4204 |
| Muse Spark 1.3 | 4202 |
| Opus 4.8 | 4102 |
| Gemini 3.1 Flash-Lite Preview | 3964 |
| Gemini 3 Pro Preview | 3893 |
| GPT-5.6 Sol | 3881 |
| Gemini 2.5 Flash Preview 04-17 | 3873 |
| Gemini 2.5 Pro Experimental 03-25 | 3871 |
| Gemini 3.5 Flash | 3845 |
| Gemini 2.5 Pro Preview 05-06 | 3836 |
| Opus 4.5 | 3797 |
| Sonnet 5 | 3713 |
| Gemini 2.0 Flash | 3659 |
| Sonnet 4.6 | 3658 |
| Opus 4.6 | 3607 |
| GPT-5 | 3498 |
| GLM-4.6V | 3464 |
| Gemini 1.5 Flash (source unspecified) | 3458 |
| GPT-4o (source unspecified) | 3426 |
| Qwen3-VL-235B-A22B-Instruct | 3356 |
| o1 (2024-12-17) | 3330 |
| Qwen2.5-VL-72B-Instruct | 3286 |
| Grok 4.7 | 3185 |
| Claude 3.7 Sonnet | 3123 |
| GPT-5.4 | 3078 |
| Grok 4.3 | 2920 |
| Claude 3.5 Sonnet (Oct 2024) | 2919 |
| Gemma 3 27B IT | 2857 |
| GPT-4o Mini | 2831 |
| Mistral Medium 3.5 | 2715 |
| Qwen3.8 27B | 2333 |