LLMs Vs. Geolocation: GPT-5 Performs Worse Than Other AI Models

Bellingcat re-ran its AI geolocation benchmark using its own photos and a 0–10 scoring scale, finding that Google AI Mode outperformed all tested models, while GPT-5 (including Thinking and Pro) delivered faster but less accurate results than prior GPT versions such as o4-mini-high. Notably, AI Mode was the only tool to correctly identify Noordwijk in a difficult test, whereas many models hallucinated or misidentified locations, and OpenAI’s changes reduced access to previously stronger models. The piece highlights rapid shifts in model capability, regional availability of AI Mode, and cautions against relying solely on LLM outputs for geolocation.