Cursor Router Assigns Coding Models Per Request, While Cost Mode Requires Retaining Grok 4.5 as a Routing Option

Cursor has launched Cursor Router for Teams and Enterprise, replacing the workflow in which users select a fixed model with per-request decisions. The router first classifies a coding task by type and complexity, then selects an executor from the models permitted by the administrator. Users can choose Cost, Balance, or Intelligence. Balance and Intelligence are still billed at the rate of the model actually selected, while administrators can configure the default mode, model allowlists or blocklists, and soft or enforced policies requiring Auto at the team or group level.

[Cursor’s technical announcement](https://cursor.com/blog/router) says the router was trained on more than 600,000 real-world requests and then A/B tested across millions of online requests, using the AFC user-satisfaction metric as its reward signal. Cursor says Intelligence mode can reduce costs by about 60% while maintaining frontier-model quality, while early enterprise customers recorded cost reductions of roughly 30% to 50%. However, the announcement does not disclose the classifier, the task-by-task model-selection matrix, confidence intervals, or test data that third parties could use to reproduce the results.

The change now covers the desktop app, Web, iOS, CLI, and SDK. A [recent TestingCatalog overview](https://www.testingcatalog.com/) confirmed the three modes and the scope of the rollout across interfaces. Notably, routing results can be hidden by default, and the Cost path requires Grok 4.5 to remain available as a cost-effective candidate. This means enterprises concerned with data residency, vendor contracts, or reproducibility cannot approve only “Cursor” itself; they must also govern the underlying model pool and routing logs. Engineering teams should require records of the actual model, version, rate, and post-failure fallback chain used for every request, and benchmark cost, edit acceptance rate, and regression errors against tasks from their own repositories instead of accepting aggregate average savings at face value.