Choose and deploy the right model
We match the model, harness and reasoning level to your work and budget, using the benchmark and your own examples, then set it up where your team works.
Which AI model is best for your geospatial work?
We benchmark 30+ AI models on real GIS and Earth observation tasks and publish every score, cost and failure. Then we build the system around your team’s data and software.
Try:
What the benchmark says now
Geospatial Intelligence Index, results as of 2 November 2026.
- Claude Opus 5.5 (xhigh) leads the Geospatial Intelligence Index at 89 of 100; GPT-6.1 Sol (xhigh) is second at 86; a gap under about 3 points is within noise.
- The best open-weights model is DeepSeek V4 Pro (max), at 79, 13th of 35.
- The cheapest model within 5 points of the leader is GPT-6 Sol (high), at $0.082 per agent task, scoring 85.
- 35 models are ranked; every score is the share of tasks passed, on the same tasks and limits.
- See how we score models (opens in a new tab)
What we do
We benchmark AI models on geospatial work and publish the results in Arena. Teams hire us to put the right model to work. AI labs hire us for the geospatial tasks and data their models learn and are measured on.
Build a spatial AI system on your data
An organisation chat, a coding workspace or a rebuilt workflow, connected to your documents, spatial data and software, and run with your team until it holds up.
Move desktop GIS work to the cloud
axis-migrate rebuilds ArcGIS and QGIS workflows as Python on the cloud you already use, and compares every output with the original.
Geospatial benchmark tasks
Evaluation tasks built on open satellite imagery and GIS data: a realistic scenario, the records an agent has to check, and a worked answer key. The same work behind Arena.
Geospatial training data
Tasks, worked solutions and step-by-step reasoning written by geospatial specialists, for teams training models on GIS and Earth observation work.
What we have shipped
Geospatial Intelligence Index
An independent benchmark of AI models on geospatial work (GIS and Earth observation), run in the agent harnesses people use, such as Claude Code and Codex. Cost per task sits beside every score.
- 74 entries
- models and harnesses, scored the same way
- 8 benchmarks
- public geospatial evaluations, under one protocol
axis-migrate
axis-migrate rebuilds an ArcGIS or QGIS workflow as Python and checks every output against the original. Use it as an open-source plugin for Claude and Codex, or as a hosted service.
- ArcGIS or QGIS → Python
- your desktop workflow, rebuilt as code
- Every output
- compared with the original
Reeve
A second brain for your farm. Reeve keeps track of your records and the rules, prices and pest bulletins around your fields, and answers on Telegram with a source for every figure. Sample farm, France only so far.
- Before you ask
- deadlines, spray days and machine alarms flagged early
- Any language
- by voice or text, on Telegram
We build around the work your team already does.
- You bring a task your team does every week, the data behind it and the software you already use.
- We pick the models with the benchmark, build the system and connect it to your tools.
- We run it with your team on real work and show how each result was checked.
- See our work
Bring us your geospatial work.
Start from a benchmark result if you like. Tell us the model and the task you were looking at, and how the work is done today.