Unveiling the Top Contender: Gemini 2.5 Pro Shines in GeoGuessr Challenges

In a quest to explore the capabilities of various language models in the popular game GeoGuessr, I recently embarked on an enlightening project that involved comparative analysis. The initiative, which can be found on GitHub, delves into how different models perform as they attempt to guess locations based solely on visual cues.

The findings have yielded some intriguing insights into the behavior of these models, which I’ve documented in a series of blog posts available here. Unsurprisingly, Google’s language models have emerged as the standout performers, likely due to their extensive integration with Google Street View, which provides invaluable contextual data for geographical guessing.

Among the contenders, the experimental Gemini 2.5 Pro has proven to be exceptionally adept, demonstrating remarkable accuracy and intuitive guesswork. Its performance is noteworthy and raises the bar for future models.

For those interested in a visual representation of the results, the leaderboard for this project ranks the models based on their performance.

Join me as we dive deeper into the fascinating world of language models and their implications for games like GeoGuessr!