GeoBench
Multimodal · 2025-03-22
GeoBench tests whether multimodal models can locate GeoGuessr imagery from a single still image, naming the country and coordinates. Its headline score is the mean 0-5000 GeoGuessr round score on the 100-location ACW (A Community World) map; Atlas imports it only for runs with no refusals, because the leaderboard averages over answered rounds.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 4404 |
| Gemini 3 Flash Preview | 4333 |
| Gemini 3.8 Flash | 4307 |
| Claude Fable 5.1 | 4303 |
| Gemini 3.5 Flash | 4288 |
| Gemini 3.1 Flash-Lite Preview | 4229 |
| Gemini 3.1 Pro Preview | 4172 |
| Gemini 3 Pro Preview | 4145 |
| Opus 4.8 | 4128 |
| Muse Spark 1.3 | 4103 |
| Gemini 2.5 Pro Preview 05-06 | 4085 |
| Gemini 2.5 Pro Experimental 03-25 | 3999 |
| GPT-5 | 3937 |
| Gemini 1.5 Flash (source unspecified) | 3934 |
| Gemini 2.0 Flash | 3897 |
| Sonnet 5 | 3878 |
| Opus 4.5 | 3877 |
| o3 | 3858 |
| Gemini 2.5 Flash Preview 04-17 | 3854 |
| GPT-5.6 Sol | 3843 |
| GPT-4.1 | 3791 |
| Gemini 2.5 Flash | 3786 |
| Sonnet 4.6 | 3762 |
| GLM-4.6V | 3730 |
| GPT-4o (source unspecified) | 3696 |
| Grok 4.7 | 3525 |
| Qwen2.5-VL-72B-Instruct | 3448 |
| GPT-4o Mini | 3401 |
| Claude 3.5 Sonnet (Oct 2024) | 3327 |
| Qwen3-VL-235B-A22B-Instruct | 3273 |
| Sonnet 4.5 | 3252 |
| Grok 4.3 | 3229 |
| Mistral Medium 3.5 | 3208 |
| Gemma 3 27B IT | 3087 |
| Llama 4 Maverick Instruct | 3069 |
| Opus 4 | 2915 |
| Grok 4 | 2719 |
| Sonnet 4 | 2468 |
| Qwen3.8 27B | 2377 |
| Pixtral 12B | 2131 |