GeoBrowse - Level 1 State (Direct): leaderboard
GeoBrowse (2026) Level 1: models compose weak visual cues in a street scene into a location. This board scores the state-level targets with the model answering directly without tools; accuracy in percent from an LLM judge, mean of three runs, higher is better.
Metric: Accuracy (%) on the state-level targets of GeoBrowse Level 1 (199 street-scene instances whose location must be composed from weak, fragmented visual cues), judged correct by an LLM judge with the Humanity's Last Exam judging prompt, mean of 3 runs, under direct inference: the model answers in a single pass without tool calls; higher is better. Source: arxiv.org. Saturation forecast: Around April 2028. 12 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 27.3 | #146 |
| 2 | Claude Opus 4.5 | 23.6 | #165 |
| 3 | Gemini 2.5 Pro | 21.8 | #280 |
| 4 | GPT-5 | 21.8 | #185 |
| 5 | Gemini 2.5 Flash | 18.2 | #466 |
| 6 | GPT-4o | 16.4 | #754 |
| 7 | Qwen 3 VL 32B | 16.4 | #624 |
| 8 | GPT-4.1 | 14.5 | #506 |
| 9 | Llama 3.2 90B | 10.9 | #1761 |
| 10 | GPT-5.2 | 9.1 | #190 |
| 11 | Qwen 2.5 VL 32B | 9.1 | #944 |
| 12 | Qwen 2.5 VL 7B | 6.3 | #1878 |
No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=geobrowse-level-1-state-direct · How It Works · Data refreshed daily, snapshot 2026-11-22.