Cursor Router Selects a Model for Each Coding Task, Balancing Quality and Inference Cost with Online Behavioral Signals

Cursor has introduced Cursor Router for Teams and Enterprise, transforming Auto mode from a system in which developers select a fixed model into one that classifies every request before sending it to a model. The router examines the query, existing context, task complexity, and domain: routine changes go to more cost-efficient models, UI work goes to models that better match the product’s preferences, and long-running or difficult problems are assigned to frontier reasoning models.

Cursor says the classifier was trained on more than 600,000 real-world requests and A/B tested across millions of online requests. Rather than relying solely on offline coding benchmarks, it estimates output satisfaction based on whether users proceed to the next task, whether they correct the agent, and the proportion of generated code that remains in the codebase over time. Training and cost evaluations also account for cache misses caused by switching models, providing a more realistic view of multi-turn agent workloads than simply comparing prices per million tokens.

Users can choose among Intelligence, Balance, and Cost modes to set their preferred tradeoff between quality and expense. Administrators can enable the router by team, restrict available modes, configure defaults, and create model allowlists or blocklists. The router is available across the desktop app, Web, iOS, CLI, and SDK, allowing the same model-selection policy to apply to both interactive development and automated agents.

Cursor claims that Auto Intelligence achieves satisfaction levels close to those of high-end models while reducing team costs by approximately 60%. Three high-usage early customers reportedly saved about 30% to 50% compared with routing all requests to Opus 4.8. These results should still be treated as internal vendor measurements: Cursor has not disclosed the accuracy of its AFC satisfaction classifier, the actual traffic share assigned to each model, or failure cases, and routing introduces the model version as an additional variable. Teams adopting the system should log the model selected for every request, along with cost, latency, and regression results, to prevent improvements in aggregate averages from masking quality degradation in specific languages or task categories.