Atlas

Benchmarks

← All benchmarks

HardGeoBench Avg Distance

Multimodal · 2025-05-08

HardGeoBench average geodesic distance in miles between a model's predicted photo location and the true location. Lower scores indicate more accurate geolocation; official leaderboard rows use pass@1 raw model calls without web search or tools.

Top models (lower is better)

ModelScore
Gemini 3.1 Pro Preview733
Claude Fable 5800
Gemini 3 Flash Preview842
Gemini 3.5 Flash886
Gemini 3 Pro Preview990
Sonnet 4.61043
o31081
Gemini 2.5 Pro Preview 06-051119
Gemini 2.0 Flash1128
GPT-5 (2025-08-07)1158
Gemini 2.5 Pro Experimental 03-251172
GPT-5.51196
Gemini 2.0 Flash-Lite1216
Opus 4.81245
GPT-51254
Opus 4.71279
GPT-5.11303
o3-pro1340
Gemini 2.5 Flash Preview (09-2025)1343
GPT-5.6 Sol1381
Gemini 3.1 Flash-Lite Preview1435
Gemini 2.5 Flash Preview 05-201444
GPT-4.11494
Opus 4.61511
GPT-5.2 (2025-12-11)1573
ChatGPT-4o Latest (source-unspecified snapshot)1578
Opus 4.51582
GPT-5 Mini1595
Gemini 1.5 Flash1763
GPT-5.41779
Kimi K2.51902
Sonnet 51920
GPT-5.6 Terra1991
Gemma 4 31B IT2065
Haiku 3.52128
O4 Mini2157
Grok 4.52185
GPT-5.6 Luna2228
Grok 4.1 Fast2329
Grok 42367
Loading Atlas data…