AI
Cursor, Ramp, and Meta race to build AI model routers—but only two are betting on their own frontier models
As developers tire of manually choosing between cheap and powerful AI models, three major companies have launched routing systems that automatically match tasks to the right model—while positioning their own AI bets for wins.
The New Stack
Cursor, the AI coding tool now backed by SpaceX, has just rolled out a model router that automatically directs coding tasks to whichever AI model is best suited to handle them. The system examines the complexity and nature of each request, then chooses between cheaper, faster models for routine work and more powerful options for genuinely challenging problems—letting developers stop thinking about benchmarks and pricing tiers.
The timing is significant. Three separate model routing systems launched within days of each other, each reflecting a different strategy for the industry's shift toward mixing and matching models rather than picking one favorite.
Why developers need routing now#
The appeal is straightforward: most engineers stick with a single model regardless of what they're doing, which means they pay frontier-grade prices for work that doesn't warrant it. Cursor's approach uses a triage system inspired by hospital emergency rooms, examining context like code complexity and request type to route appropriately. The company claims early customers saw cost savings of 30–50% versus sending everything through expensive models, without sacrificing output quality.
Three optimization modes let developers lean into either speed and cost or raw capability, depending on their immediate needs.
David Pan, Cursor's field CTO, framed the problem bluntly: "We briefly went insane and decided every software engineer should also become an expert in model benchmarks, thinking levels, and cache hit rates." That sentiment has resonated—engineers already juggle this trade-off manually, defaulting to cheap models for daily work and splurging on expensive ones only for serious tasks.
Cursor's own model stack#
The launch reflects Cursor's broader push to control its AI infrastructure. The company released Composer 2.5, its in-house coding model, in May. Now with SpaceX's backing—valued at $1.5 trillion post-IPO—Cursor jointly developed Grok 4.5, a new frontier model priced at $2 per million input tokens and $6 per million output tokens.
Router solves a practical problem: Cursor now has cheap internal models and expensive external ones. Rather than always pushing requests to its own offerings, Router intelligently routes each task to whichever model actually fits—maximizing quality while keeping costs down and justifying use of Cursor's own models where they make sense.
