Using artificial intelligence can get expensive very quickly, especially when companies are sending huge numbers of requests to models from OpenAI, Anthropic, and others. Ramp thinks businesses are wasting money by using more powerful AI models than they actually need.
The financial technology company has launched Router.com, an AI model router designed to automatically send each request to the lowest-cost model capable of delivering the required level of performance. Ramp says customers already using Router have reduced inference costs by 40 percent on average.
Router essentially gives developers a single API endpoint for accessing a collection of major AI models. Instead of an application being tied directly to one model, Router can decide which model should handle a particular request based on cost and performance requirements.
At launch, developers can access models from OpenAI, Anthropic, and SpaceXAI, along with open models from Nvidia, Kimi, DeepSeek, GLM, and Qwen. Ramp says Google’s Gemini is coming soon.
The service also works with AI infrastructure providers such as Fireworks AI. Support for Google, AWS, Together AI, Baseten, Crusoe, and additional providers is planned.
Of course, simply putting several models behind one API isn’t particularly novel. The more interesting part of Ramp’s approach is how aggressively it is trying to optimize which model gets used.
Router includes more than 100 optimizations covering model selection, caching, compression, timing, and request handling. It can also automatically fall back to another provider when one fails.
Ramp says it continuously evaluates new models using Ramp SWE-Bench, an internal benchmark built around actual engineering work performed at the company. Rather than relying entirely on public AI leaderboards, Ramp uses those production tasks to determine which models provide the best combination of performance and price.
As new models and routing techniques perform well in those tests, Ramp can incorporate them into Router’s defaults. Developers can use Ramp’s routing strategy or configure their own.
Interestingly, Router isn’t something Ramp cooked up solely as a new product. The company says it originally built the technology for its own use about three years ago.
According to Ramp, its internal routing system reduced inference costs by approximately 30 percent while maintaining the same output and delivering more than 99.9 percent reliability across production traffic.
The timing certainly makes sense. Ramp claims AI spending has increased 20.7 times since June 2025, based on data from its own Ramp AI Index. As businesses use AI for more tasks, choosing an unnecessarily expensive model thousands or millions of times can turn into a surprisingly large bill.
There is an obvious tradeoff, however. Automatically choosing a cheaper model only saves money if the selected model actually produces acceptable results. A 40 percent reduction in inference costs won’t mean much if developers end up dealing with lower-quality output or inconsistent behavior.
Ramp believes its benchmarking and routing system can avoid that problem by matching models to individual workloads rather than assuming the biggest or newest model is always the right choice.
Router.com is available now and Ramp says routing will be free through the end of 2026. New users also receive $26 in credits, while model usage itself is billed at the applicable token prices.
For companies already juggling several AI providers, the pitch is pretty straightforward. Instead of developers constantly deciding whether a particular request really needs an expensive frontier model, Router can make that decision automatically.
If Ramp’s claimed 40 percent savings hold up across a wider range of real-world workloads, that could become increasingly attractive as corporate AI bills continue climbing.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz