Coordinates and targets
Initial location state: WGS84 latitude/longitude only. Identical fixed definitions and three independent best-estimate questions. Every target requires one allowed class; no not_sure choice.
GEOGRAPHY FROM COORDINATES
Where on Earth?
A geography benchmark for AI models
192 coordinates · 4 models · 19 settings
Contiguous United States
GPT-6 Luna, GPT-5.6 Terra, GPT-6.1 Sol and Jev 1.13.0 classify geography from latitude and longitude. The same fixed terrain, land-cover and water definitions; coordinate-only initial location input.
THE ONTOLOGY
Three independent targets.
Hover, focus or tap a class for its exact definition.
THE BENCHMARK
Model agreement with the references.
| Model | Terrain | Land cover | Water | Response completion |
|---|
Most-common-class baseline: the most frequent reference class in this sample.
Model performance
Source agreement across reasoning and search settings. Select a marker to update the table and maps.
3,644 / 3,648 valid responses
Four GPT-6.1 Sol high/search-on provider errors; no retries.
Search offSearch onJev 1.13.0: one baseline, effort/search N/A
Each chart uses its own zoomed Y axis; tick labels show the range. Agreement uses 180 land references; provider errors count as misses. Water: 178 outside mapped water / 2 lakes, with a 98.9% constant-outside baseline.
INDIVIDUAL COORDINATES
The same locations, across four models.
Hover or focus a location for its prediction, reference and condition. Tap or select it to compare all three targets. Circles are land coordinates; diamonds are separate offshore controls.
Matches referenceDiffers from referenceOffshore terrain/cover unavailableProvider error◇ Ocean control
All 180 land coordinates are scored for every target. The 12 offshore controls test water only. These are sampled answers, not a continuous reconstructed surface; Alaska/Hawaii are excluded.
METHODOLOGY
One shared sample, one controlled comparison.
Sampling and references
140 interior and 40 boundary land points, selected geometrically from a 485-point, 125 km EPSG:5070 equal-area frame. Twelve offshore ocean controls are separate. Alaska/Hawaii are excluded.
Reasoning and search
Jev 1.13.0 baseline plus GPT-6 Luna, GPT-5.6 Terra and GPT-6.1 Sol × low/medium/high effort × search off/on: 19 settings. Only effort and access policy vary. Search permits one optional public text query; direct reference/GIS answers are excluded.
Scoring and responses
Correct / eligible source references; missing predictions count as misses. All 180 land points are scored for all three targets; the 12 ocean controls are water-only. 3,644 valid answers / 3,648 conditions; four provider errors, no retries.
GPT confidence is an uncalibrated self-report; Jev 1.13.0 has no equivalent confidence, reasoning or search interface. Source agreement and single responses per setting do not establish national accuracy or statistical significance.
Reference sources: Named Landforms of the World v3 (CC BY 4.0), USGS Annual NLCD 2024, OpenFreeMap / OpenMapTiles and OpenStreetMap contributors (ODbL). Census/GSHHG display outlines.