Atlas

Benchmarks

← All benchmarks

HardGeoBench <=10mi

Multimodal · 2025-05-08

HardGeoBench percentage of private-photo geolocation predictions within 10 miles of the true location. Scores are pass@1 percentages from API default settings with a simple prompt, no sampling, and no web search or tools.

Top models (higher is better)

ModelScore
Gemini 3.5 Flash45.0
Claude Fable 541.3
Gemini 3 Flash Preview38.8
Gemini 3.1 Pro Preview38.8
Gemini 2.5 Pro Preview 06-0537.5
Gemini 2.5 Pro Experimental 03-2535.0
o3-pro35.0
Gemini 3 Pro Preview35.0
o333.8
GPT-5.133.8
Opus 4.632.5
ChatGPT-4o Latest (source-unspecified snapshot)31.3
GPT-531.3
GPT-5 (2025-08-07)31.3
GPT-4.130.0
Opus 4.530.0
Kimi K2.530.0
GPT-5.6 Sol30.0
GPT-5.528.8
Gemini 2.5 Flash Preview 05-2027.5
Gemini 2.5 Flash Preview (09-2025)27.5
Sonnet 4.627.5
Gemini 3.1 Flash-Lite Preview27.5
GPT-5.427.5
Opus 4.727.5
Opus 4.827.5
Gemini 2.0 Flash26.3
GPT-5.6 Terra25.0
GPT-5.2 (2025-12-11)23.8
Gemini 2.0 Flash-Lite22.5
O4 Mini21.3
GPT-5 Mini21.3
Sonnet 4.521.3
Haiku 3.520.0
Sonnet 520.0
Grok 418.8
GPT-5.6 Luna18.8
Grok 4.517.5
Gemini 1.5 Flash16.3
Gemma 4 31B IT15.0
Loading Atlas data…