A week using M3 as Cursor’s backend for real coding work
M3 was connected as the inside Cursor and used for a full week of real development work. The work included bug fixes, a large refactor in a hard-to-touch service, and writing tests after the code was already changed. Setup was simple because the API was : change the base URL, add the key, and keep the rest the same.
Cursor and Claude Code worked without a shim, proxy, or long setup detour. The strongest result was in repeated AI work loops: run a command, read the error, fix the code, and run it again. The model kept the flow coherent across and did not lose shell state as often as cheaper backends usually do.
Multi-file edits also worked well, especially during a rename refactor where call sites across files stayed in view. The large helped it keep references from a whole module available deeper into the session.
Key points
- M3 was used as Cursor’s for a week of real coding work.
- Setup only required changing the base URL and adding an API key.
- Cursor and Claude Code worked without a shim or proxy.
- It handled command-error-fix-rerun loops better than many cheaper backends.
- It tracked multi-file rename work and benefited from a large .