Anysphere, the company behind the popular Cursor code editor, has launched a tool called Cursor Router. For each request, it automatically determines which model is best suited to the task and sends it there. You get results on par with the most powerful models, but only pay premium prices when the work requires it. According to Cursor's figures, this results in 30 to 60 percent lower costs at comparable quality.
Using one model for everything makes work more expensive
Cursor noticed in its own data that roughly six out of ten developers choose one model and use it all day. As a result, even trivial changes go through the most expensive model available, and the AI bill grows faster than the quality of the work delivered. With Router, each request is classified first, and only then is a model selected to handle it. Simple tasks go to cheaper and faster models, interface work goes to the model with the best eye for visuals, and complex, long-running tasks are assigned to the most powerful models built for demanding reasoning.
How Router knows where to send a request
Router includes a classifier that makes a decision before any model starts running. It considers the query itself, the available context, the task's complexity, and its domain—in other words, whether it involves server logic or an interface. It also draws on its own knowledge of what each model does well.
Cursor trained the classifier on more than 600,000 requests and evaluated it in online A/B tests across millions of real queries. It used user satisfaction as the training reward. When a developer moves on to the next feature after receiving a response, the system takes that as approval. When they start correcting the agent, it reads that as a signal that something went wrong. The company designed the tool from the outset for a world where new models are released almost every month, so it can simply add more to the lineup without having to rebuild the entire system.
One detail makes a big difference in practice. Switching from one model to another in the middle of a conversation can break the cache—the temporarily stored context—adding both cost and latency. Cursor says it accounts for this in training and measurement, and that the reported savings include this cost. A router that ignored it could quietly consume most of the savings it touts.
Three modes based on what you need
Router offers three settings that can be selected through the Auto option in the model picker. Intelligence mode aims for maximum quality comparable to the most expensive models. Balance mode combines solid quality with a lower price. Cost mode keeps expenses as low as possible while trying to deliver the highest reasonable quality. The decision remains with the user. An individual can switch to Intelligence for a difficult refactoring task, while a company can set Balance as the default for everyone to keep the bill under control.
How much you can actually save
These figures were published by Cursor and have not been independently verified. Auto Intelligence came close to the Fable model in satisfaction while costing teams roughly 60 percent less. Compared with Opus 4.8, it increased satisfaction by about 15 percent for nearly the same price. Auto Balance also outperformed Opus 4.8 while costing 36 percent less. Compared with GPT-5.6 Sol, it delivered comparable satisfaction at a lower cost. Prices per commit:
- Auto Balance: $4.63
- Auto Intelligence: $6.76
- Opus 4.8: $7.34
- Fable 5: $12.69
GPT-5.6 Sol matched Intelligence mode on price, but users were less satisfied with its output. This difference is precisely the practical argument for routing. Router leaves difficult tasks to the most capable models and moves routine work off the expensive rate. The results were similar for real customers. During two weeks of trial access, three large companies with a combined total of thousands of users saved 30 to 50 percent on automatically routed requests compared with sending everything through Opus 4.8.
What to watch out for
Automatic routing is a genuine lever for reducing costs, but it comes with caveats. The first is visibility into which model did the work. When a machine selects the model, you do not always know which one wrote your code. Some developers have already asked for the ability to see which model was used for each task so they can review the output accordingly. The second issue is vendor lock-in, because Router optimizes across Cursor's models and according to its own metric. Third, the savings figures are Cursor's own data, which have not been independently verified. It therefore makes sense to test it first in your own environment.
Router and its availability
Router is available on Teams and Enterprise plans, on desktop, the web, iOS, via the command line, and through the SDK. Administrators can choose whether to enable it for the entire team or only certain groups, specify the available modes, set a default, and allow or block specific models in Router.
