Developer Simon Willison documented a GPT-6 Astra (Max) task on September 12, 2026, that converted a home address into 5K and 10K running loops. The task ran inside ChatGPT Work for 27 minutes and returned an embedded map alongside downloadable GPX and GeoJSON files.
Willison's account describes a multi-stage geospatial workflow rather than conversational route advice. GPT-6 Astra used Nominatim to locate the address, retrieved nearby OpenStreetMap roads and trails through Overpass, and calculated loop routes locally. One displayed route measured 5.1 km, close to the requested 5K distance.
The workflow matters because each stage would usually require a separate GIS tool or integration. Geocoding converts an address into coordinates, Overpass retrieves relevant OpenStreetMap features, and local computation turns that network into distance-constrained loops. Visualization and file export then make the result usable by a person or another application.
It worked like a compact GIS workbench controlled through one task request. The practical value came from coordinating data retrieval, calculation and presentation, not from producing plausible directions in prose.
The visualization was self-contained enough to embed in the ChatGPT interface. Its HTML stored the full route and map geometry in a JSON element, then used D3 version 7.9.0 from an allowed content-delivery network to render the map. The output also credited OpenStreetMap contributors.
The supplied account does not include a parameter count, model architecture, training details or context-window size. It provides no license, API availability or pricing information. The only demonstrated access point is GPT-6 Astra (Max) within ChatGPT Work, so the report does not establish whether developers can call the same model or tools through an API.
The account also contains no benchmark result or controlled comparison with an earlier model. It documents one completed task, not evidence that Astra sets a new state of the art for geospatial reasoning. Route quality was not scored against specialist running applications, and the account does not report tests for elevation, road crossings, surface type or pedestrian access.
Those gaps limit what can be concluded from the result. The demonstration shows that the model completed the described workflow, but it does not establish that the generated routes are safe, legal or suitable for a particular runner. Geometric distance alone is not sufficient validation.
The workflow is most relevant to builders prototyping location-based tools without first connecting every geospatial component by hand. Potential applications include running-route experiments, local trail discovery and other services that need geometry in portable formats. GPX supports transfer into compatible navigation and fitness tools, while GeoJSON provides structured geographic data for further processing.
A production service would still need explicit checks for pedestrian restrictions, crossings, terrain and current conditions. Evaluation should cover the complete artifact rather than only the final map, including route validity, distance accuracy and export usability before an agent-generated loop reaches a consumer product.
