GeoGuessr Agent 🌍

Solving GeoGuessr games using VLMs to 90%+ accuracy.

❓ What is GeoGuessr: the game puts you into Google Street View and challenges you to guess the location. Play for free here.

How does it work?

🤖 Uses Playwright to play the game and move around.
🎯 Automatically spots and zooms in on interesting stuff (signs, landmarks, etc.) to get a better view.
🔍 Location detection using GPT-4o and O1 vision models

The Results 📊

Played 5 games (25 rounds) of the Classic Map with no-move (but allow pan and zoom).

o1 performs insanely well with 91%+ accuracy! 4o is significantly cheaper, faster, and performs almost as well.

Model	Score %	Avg Score/Game (/25,000)	Best Game (/25,000)	Worst Game (/25,000)	Median Guess (km)	Best Guess (km)	Worst Guess (km)
GPT-4o	89.9%	22,464.2	24,530	20,561	160.4	0.2	626.7
o1	91.8%	22,943.6	24,777	21,494	76.6	0.2	523.3

Running

Copy .env.example and set the keys.

The Geoguessr token is found in the browser's cookie.

uv run main.py

Implementation Notes

• gpt-4o required prompt tweaking to work well. o1 was smart enough to work with a very basic prompt.
• Performance drops on the A Community World (ACW) map to ~80% for o1. Why?
• How much does the zoom-in feature improve results?