Every company running artificial intelligence at scale eventually hits the same problem: a simple customer question gets routed to the same, expensive, frontier-grade model as a genuinely hard one. That’s a waste of money — and an avoidable one. Routing software fixes it by evaluating each request before it happens and sending it to whichever AI model is suited to the job, saving the expensive model for the tasks that need it. Done well, intelligent routing can cut a company’s AI bill by 30% to 50%, with some workloads seeing savings as high as 85%, according to Mindstudio.
In a single week this July, four companies moved to build, buy or open their own version of AI model routing.
Cursor, Ramp and Meta Are Chasing the Same Layer
Cursor Router, launched Tuesday (July 21) evaluates each coding request first and directs it to the model that actually fits the task. About 60% of Cursor’s users had been defaulting to a single AI model regardless of what they were working on, AlphaSignal reported. Cursor says its Router delivers the same quality of output at 60% lower cost in internal testing, with early enterprise customers reporting savings of 30% to 50%, according to the report.
Spend management company Ramp spent three years building a routing tool for internal use, running it across more than 100 of its own AI features, running it across more than 100 of its own AI features. It has cut internal AI costs by 30%, the company said in its launch announcement. Rather than keep that edge to itself, Ramp opened the tool to any business for free, charging only for the underlying AI usage. Business AI spending among Ramp’s customers rose nearly tenfold between January 2025 and April 2026, according to figures the company shared publicly on social media.
Meta is building its own internal routing system, reportedly called Switchboard, that sends simple requests to cheaper models and reserves expensive ones for harder tasks, The Information reported. Meta has not said whether it plans to sell the tool externally and it has not confirmed the report. That alone is telling. Even a company with Meta’s resources is building custom infrastructure just to keep the AI bill down.
Then, in the same stretch of days, came the move that made the pattern impossible to ignore. Stripe was reportedly negotiating to acquire OpenRouter, a startup already running AI model routing at massive scale, for roughly $10 billion, PYMNTS reported.
Owning the Routing Decision Means Owning the Bill
For a payments company like Stripe, the routing layer is valuable for a reason that has little to do with AI cost savings. It sits one step upstream of a problem FinTech companies already spend heavily trying to solve: billing AI usage accurately. Unpredictable AI pricing is the primary barrier to adoption for 46% of IT leaders, according to Stripe’s own research. AI infrastructure costs have climbed from roughly 10% of a typical AI company’s budget to as much as 35% to 40% as usage scales, industry analysis from billing platform Flexprice found. And the price itself keeps shifting. Ninety-two percent of AI companies that started with usage-based pricing have changed the model at least once, according to 2025 research from Metronome, the billing company Stripe acquired in January, as cited by Flexprice.
Long before an invoice gets generated, something has already decided which model answers a request and what it will cost. Whoever owns that decision owns the first move in every transaction that follows. A simple question still doesn’t cost the same to answer as a hard one, and increasingly, neither does owning the software that decides which is which.
