GISAgentBench: leaderboard
GISAgentBench (University of Florida, 2026) tests LLM agents on 349 multi-step GIS tasks drawn from GIS Stack Exchange questions and recast onto real public data in six areas of interest, working through a fixed harness of 128 GIS APIs. Strict task success rate in percent: an output file counts only when every structural check passes and every value matches the exact ground-truth output within tolerance, with no LLM judge; higher is better.
Metric: Strict task success rate (%; share of the 349 tasks whose output file matches the exact ground-truth output within the task's declared tolerance under deterministic structural and value checks, no LLM judge; LangChain ReAct agent over a fixed harness of 128 GIS APIs backed by QGIS, GDAL and GeoPandas; 50 tool calls per task, temperature 0, one run per task, July 2026). Source: arxiv.org. Saturation forecast: Around June 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 32.7 | #94 |
| 2 | Claude Opus 4.6 | 29.5 | #102 |
| 3 | DeepSeek V4 Pro | 27.5 | #188 |
| 4 | GPT-5.4 | 22.9 | #123 |
| 5 | Qwen 3.6 27B | 17.8 | #341 |
| 6 | GPT-OSS-120B | 12.3 | #700 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gisagentbench · How It Works · Data refreshed daily, snapshot 2026-11-13.