Cursor launches Router: a classifier that distributes each coding task among models to cut AI spending

Cursor has unveiled Cursor Router, an intelligent model-routing system designed for teams and enterprise clients. The core idea is simple: instead of each developer manually choosing a single model as their "workhorse" for all tasks, Router analyzes each request before executing it and decides which model is best suited to solve it, aiming to deliver frontier-level intelligence without always paying the price of a frontier model.

The problem the company says it is tackling is concrete: according to its own data, around 60% of the developers who use Cursor set a single model as the default option for all their work. This means routine tasks (small changes, interface tweaks, trivial fixes) are paid for at the same price as complex long-reasoning problems, causing AI spending to grow much faster than the actual quality of the result obtained.

At its core, Cursor Router is a classifier trained on more than 600,000 real usage requests, and subsequently evaluated through online A/B tests over millions of requests in production. The system evaluates each request based on the query itself, the context, the complexity of the task and the domain it belongs to, cross-referencing that information with what the company has learned about the strengths of each available model. As they describe it, simple work is routed to the most cost-efficient models, interface updates go to the model with the best visual "taste," and complex, long-horizon problems are reserved for the most powerful reasoning models. A relevant technical detail is that the classifier is trained and evaluated while being aware of the cost of cache misses: switching models mid-conversation has an additional cost, and Cursor states that it incorporates that cost both into training and into the savings figures it reports.

To measure the system's effectiveness, Cursor deliberately opted for large-scale online A/B tests instead of offline evaluations, arguing that the latter are limited in size, are far removed from real usage and are hard to reduce to a success rubric, in addition to not capturing the cost of model switches between the turns of a conversation. The quality metrics it says it has used for nine months to evaluate each model release or system improvement are two: user satisfaction (interpreting moving on to the next task as a strong positive signal, while correcting the agent is a strong negative signal) and the code retention rate—that is, how much of the code generated by the agent remains present in the repository over time.

As for results, during a two-week early-access period with a group of enterprise clients, Cursor compared what they actually paid against what the same activity would have cost if everything had been processed at Opus 4.8's API rates. Three high-volume accounts, with thousands of users each, saved between 30% and 50% on the automatically routed requests, with no drop in perceived quality. In online A/B tests over millions of requests, the company reports savings of up to 60% while maintaining frontier-level performance.

The system offers three configurable modes that move along what Cursor calls the "Pareto frontier" between cost and intelligence: Intelligence (frontier quality, on par with the most powerful and expensive models), Balance (high quality, similar to the frontier models most users prefer for daily use) and Cost (good quality, prioritizing token savings). According to the figures the company provides, Auto Intelligence mode sits close to Fable in user satisfaction at a roughly 60% lower cost for teams, and beats Opus 4.8 in satisfaction by around 15% at a practically equivalent price. Auto Balance mode, for its part, comes in above Opus 4.8 in satisfaction at a cost around 36% lower, and against GPT-5.6 Sol it offers comparable satisfaction at lower spend.

Beyond cost per request, Cursor also analyzed cost per commit—that is, per each delivery of code effectively completed. According to its data, a commit resolved with Router's Intelligence mode costs $6.76 and with Balance mode $4.63. GPT-5.6 Sol matches the cost of Intelligence mode but with lower user satisfaction with the result, while Fable 5 and Opus 4.8 produce commits at a cost higher than Cursor Router's: $12.69 and $7.34 respectively. For Cursor, this is the practical proof that automatic routing keeps hard tasks on the most capable models while steering routine work away from frontier prices.

On the administrative side, organizations retain control over how the tool is deployed: they can enable it by team or group, decide which modes each user can choose, set a default mode and allow or block specific models. Cursor Router is already available for the Teams and Enterprise plans on desktop, web, iOS, CLI and its SDK.

The company frames this launch as one more piece of its overall strategy for efficiency in token consumption, not as an isolated measure. It mentions in particular "dynamic tool calling," a mechanism by which most native tool descriptions are no longer loaded into every prompt: the model queries them the first time it needs them, following the same pattern already used for MCP servers. This keeps frequently used tools, such as reading or editing files, always available, while occasionally used ones only enter the prompt when the agent actually invokes them. Cursor also notes that it continues to expand the set of available models—citing the addition of Grok 4.5 for harder and more costly tasks—and that its own Composer model keeps improving in everyday use, so that lower-cost turns stay close to frontier quality without paying frontier prices.