Cursor has launched Cursor Router, an intelligent model-routing system intended to help engineering teams balance coding quality against model cost. The router now powers Auto mode and is available to Teams and Enterprise customers across Cursor's desktop, web, iOS, command-line and software development kit products.
The system classifies a request before a model runs, using the query, available context, task complexity and domain. It then directs the work to a model Cursor considers appropriate: lower-cost models for routine tasks, specialist choices for interface work and frontier reasoning models for more difficult or longer-running jobs.
Three optimisation modes
Users can select Intelligence, Balance or Cost modes. Intelligence aims for the strongest available quality, Balance targets a middle ground, and Cost prioritises lower spending while retaining what Cursor describes as useful capability. Teams are billed at the rate of the model selected for a request, so routing changes the mix of model charges rather than introducing a single flat rate.
Administrators can enable the router for individual teams or groups, limit which modes are available, choose a default and allow or block underlying models. They can also decide whether the selected model is shown to users and apply soft or hard rules encouraging a standard Auto configuration.
These controls matter for organisations that want the convenience of automated selection without surrendering procurement, security or model-governance decisions. A company can, for example, block a provider it has not approved even if Cursor's classifier would otherwise choose that model for a task.
Cursor's performance claims
Cursor says it trained the classifier on more than 600,000 live requests and evaluated it in online tests covering millions of requests. The company reports that its Intelligence setting achieved frontier-level user satisfaction at substantially lower cost in its tests, while selected early-access enterprise accounts saved between 30 and 50 per cent compared with sending all traffic to one premium model.
Cursor also reports lower cost per commit in its internal analysis. These are company-provided results based on Cursor's own production signals, including user satisfaction and the proportion of generated code retained in a codebase. They are not an independent benchmark, and the savings available to a particular team will depend on its workload, model allow-list, conversation patterns, caching and the prices of models available at the time.
The router is cache-aware in training and evaluation, according to Cursor. That is relevant because changing models during a conversation can create cache misses and erase part of the apparent saving. Cursor says the reported production comparisons include this effect, but it has not published enough underlying data for outside researchers to reproduce the analysis.
What changes for engineering teams
Automatic routing reduces the need for every developer to understand the strengths, limits and current prices of a rapidly changing model catalogue. It may also help finance and platform teams manage spend without forcing all coding work onto one model. The trade-off is less predictability about which model handles an individual request unless administrators choose to expose that information.
Model choice can also affect data-handling terms, regional availability, output behaviour and debugging. Organisations adopting Cursor Router should review how selected providers process prompts and code, confirm that approved models match internal policy, and monitor quality by task type rather than relying only on an average cost figure.
Cursor says the router can be updated as new models arrive. Grok 4.5 is among the routing options mentioned at launch, while Cursor's own Composer models remain part of the lower-cost path. The company is also working on token-efficiency changes such as loading less-used tool descriptions only when an agent needs them.
For Teams customers, Auto is enabled by default. Enterprise administrators can activate it from the dashboard. The launch makes routing a formal product feature rather than a manual model-picker decision, but organisations remain responsible for configuring the quality, cost and provider boundaries appropriate to their code and risk profile.