ChatGPT Work with GPT-6 Astra generates running routes using OpenStreetMap data

Simon Willison reports that a ChatGPT Work session with GPT-6 Astra (Max) generated 5K and 10K running routes from a user’s home address using OpenStreetMap (OSM) data.

The agent used Nominatim to geocode the address and Overpass to fetch local roads and trails, then computed loops locally before producing an embedded map visualization, a downloadable GPX file, and a GeoJSON file.

The embedded map was rendered via a 'visualize' skill that created an HTML file embedding the route geometry and D3.js from an allow-listed CDN (cdnjs.cloudflare.com, cdn.jsdelivr.net, unpkg.com, and others).

The author notes that the exact Python code executed was not visible in the ChatGPT UI and could not be retrieved after the thread was compacted, calling out lack of transparency and compaction as anti-features that hinder reproducibility.

Willison argues that LLM systems using compaction should preserve pre-compaction text and make it available via agent tool calls to mitigate this issue.