AGL1K: leaderboard
AGL1K (Beihang University, 2026) asks audio-language models where on Earth a crowd-sourced Aporee field recording was made, over 1,444 clips from more than 70 countries and territories selected for audio localizability and balanced between clips with and without speech. The Geo-score averages 5000 x exp(-d/1492.7) over clips for a Haversine error of d km (0-5000, higher is better), with a refusal counted as a 10,000 km error.
Metric: Geo-score (0-5000). Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 3031.82 | #135 |
| 2 | Gemini 2.5 Pro | 2826.95 | #262 |
| 3 | Gemini 2.0 Flash | 2535.49 | #702 |
| 4 | Gemini 2.0 Flash (Thinking) | 2494.46 | #702 (Gemini 2.0 Flash) |
| 5 | Gemini 2.0 Flash Lite | 2417.14 | #1032 |
| 6 | Gemini 2.5 Flash | 2256.22 | #452 |
| 7 | Gemini 2.5 Flash Lite | 1687.97 | #958 |
| 8 | Qwen2.5-Omni-7B | 1317.74 | #1797 |
| 9 | Gemma 3n E4B (IT) | 1161.54 | #2220 |
| 10 | Phi-4 Multimodal Instruct | 826.88 | #3914 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agl1k · How It Works · Data refreshed daily, snapshot 2026-11-12.