GeoBench
Identifying where a photo was taken.
As of 19 Sept 2026, Gemini 3 Flash leads GeoBench on BenchLeader with 88.0%, ahead of Gemini 2.5 Pro at 86.0%, across 27 model configurations with a published result.
- Published by
- GeoBenchdata via Epoch AI Benchmarking Hub
- Category
- Multimodal
- Index weight
- Reference only
- Models
- 27
- Data as of
- 19 Sept 2026
CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.
What the test looks like
Given a street-level photo, the model must name the country (and finer levels) where it was taken.
How it is scored
Percent of countries identified, as published by GeoBench.
What to keep in mind
Vision models only; a test of world knowledge through images.
- 1Gemini 3 Flash88.0%
- 2Gemini 2.5 Pro86.0%
- 3Gemini 3 Pro84.0%
- 4GPT-5 (medium)81.0%
- 5o1 (medium)80.0%
- 6Gemini 2.0 Flash77.0%
- 7Gemini 2.5 Flash76.0%
- 8Gemini 1.5 Flash 00276.0%
- 9Claude Opus 4.575.0%
- 10o3 (medium)74.0%
- 11GPT-4.172.0%
- 12GPT-4o71.0%
- 13Claude 3.7 Sonnet68.0%
- 14o4-mini (medium)64.0%
- 15o4-mini (high)64.0%
27 of 27
| # | ||||
|---|---|---|---|---|
| 1 |  Gemini 3 FlashGoogle | 88.0% | 57.7 | 2025-12-17 |
| 2 | Gemini 2.5 ProGoogle | 86.0% | 54.5 | 2025-05-06 |
| 3 | Gemini 3 ProGoogle | 84.0% | 61.1 | 2025-11-18 |
| 4 |  GPT-5mediumOpenAI | 81.0% | 58.5 | 2025-08-07 |
| 5 | o1mediumOpenAI | 80.0% | 48.4 | 2024-12-17 |
| 6 | Gemini 2.0 FlashGoogle | 77.0% | 43.7 | 2025-02-05 |
| 7 | Gemini 2.5 FlashGoogle | 76.0% | 52.5 | 2025-05-20 |
| 8 | Gemini 1.5 Flash 002Google | 76.0% | 40.8 | 2024-09-24 |
| 9 |  Claude Opus 4.5Anthropic | 75.0% | 58.2 | 2025-11-24 |
| 10 | o3mediumOpenAI | 74.0% | 53.7 | 2025-04-16 |
| 11 | GPT-4.1OpenAI | 72.0% | 49.0 | 2025-04-14 |
| 12 | GPT-4oOpenAI | 71.0% | 44.3 | 2024-11-20 |
| 13 | Claude 3.7 SonnetAnthropic | 68.0% | 50.2 | 2025-02-24 |
| 14 | o4-minimediumOpenAI | 64.0% | 52.4 | 2025-04-16 |
| 15 | o4-minihighOpenAI | 64.0% | 50.8 | 2025-04-16 |
| 16 | GPT-4o miniOpenAI | 64.0% | 36.2 | 2024-07-18 |
| 17 | Claude 3.5 SonnetAnthropic | 62.0% | 46.6 | 2024-10-22 |
| 18 |  Qwen2.5-VL 72B InstructAlibabaopen ↗ | 62.0% | – | 2024-09-19 |
| 19 | o3highOpenAI | 60.0% | 53.0 | 2025-04-16 |
| 20 |  Llama 4 MaverickMetaopen | 52.0% | 43.4 | 2025-04-05 |
| 21 | Gemma 3 27BGoogleopen | 52.0% | 39.5 | 2025-03-12 |
| 22 | Llama 3.2 90BMetaopen ↗ | 52.0% | 37.6 | 2024-09-24 |
| 23 | Claude Opus 4Anthropic | 49.0% | 52.9 | 2025-05-22 |
| 24 |  Grok 4xAI | 45.0% | 58.5 | 2025-07-09 |
| 25 | Claude Sonnet 4Anthropic | 37.0% | 49.7 | 2025-05-22 |
| 26 |  Pixtral 12BMistral AIopen | 37.0% | – | 2024-09-17 |
| 27 | Claude 3.5 HaikuAnthropic | 34.0% | 40.3 | 2024-10-22 |
Cite as: BenchLeader, “GeoBench leaderboard”, https://www.benchleader.com/benchmarks/geobench, data as of 19 Sept 2026.
GeoBench: questions
- What does GeoBench measure?
- Given a street-level photo, the model must name the country (and finer levels) where it was taken. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads GeoBench?
- Gemini 3 Flash leads GeoBench with 88.0% as of 19 Sept 2026, ahead of Gemini 2.5 Pro at 86.0%.
- How many models have GeoBench results?
- 27 model configurations have a GeoBench result on BenchLeader, all taken from GeoBench via Epoch AI Benchmarking Hub.
- Who runs GeoBench and how often is it updated?
- GeoBench is published by GeoBench. BenchLeader re-reads the published results every morning and records the date each result was published.
- Does GeoBench count toward the BenchLeader Index?
- No. GeoBench is shown for reference but left out of the composite index.