AI plays GeoGuessr.

21 models · the same 50 places · one image or up to eight exploration actions.

Play beside the modelsOpen GeoGuessr challengeRead Gemini 4 Argon's reasoning, round by round

Where they guessed

Choose a location and model. The diamond marks the real location.

Each guess, beside what the model saw

Choose a location to see the model guesses. Click a model to show its saved images on the right.

Drag to pan. Use + / − or Ctrl + scroll to zoom. ◇ marks the true starting location.

Saved Street View image

Model’s explanation

Click the image to open it at full size. Exploration images differ between models; the answer stays at the original starting location.

World outline: Natural Earth, public domain. Images are the original saved model inputs.

How far they missed

50 locations per model.
Lower distances are better. Distance summaries use valid guesses; failed rounds count as zero points.
ModelValid guessesTypical miss
median, km
Average miss
km
Within 250 kmAverage pointsWorst miss
km

Closest to farthest

One dot per guess. Each model is sorted independently. Select a dot to inspect it.

Select a dot to see its location and saved explanation.

Did looking around help?

One-image median Exploration-allowed median. Smaller is closer.

Which locations got another look?

One square per location. Color means at least one exploration action.

Setup and model settings

The same 50 locations, without exploration or with up to eight actions. Models choose actions and coordinates; the harness drives the website. Screenshots hide the guess map and score hints.

VersionRequested IDEffortAnswer format