Ramp launches Router.com to route AI model traffic

Ramp has launched Router.com, an application programming interface for accessing, testing and routing artificial intelligence models. The product is now available beyond Ramp's own customer base.

The launch marks Ramp's move to commercialise internal infrastructure it has used across its products to manage model selection, monitor request costs and shift traffic between providers. The system has been in use for three years on production workloads and now handles Ramp's live AI traffic.

Router.com is aimed at developers navigating a fast-moving market in which model releases, price changes and new service tiers can quickly alter an application's economics. Instead of forcing teams to rebuild integrations each time they want to test a new model, the service offers a single API that keeps the application interface unchanged while comparisons happen behind the scenes.

That comparison process is central to the product's pitch. Teams can evaluate models on output quality, time to first token, total latency, reliability and cost before deciding whether to move production traffic.

Routing options

The service includes several routing strategies designed to match different cost and performance trade-offs. One option, called Flex tier, checks whether discounted service tiers for the same model deliver latency comparable to standard tiers and directs requests to the lower-priced option when they do. If latency worsens, the request stays on the standard tier.

Another feature, shadow models, sends a sample of real requests to a candidate model while the existing model continues to serve users. This lets teams compare cost and latency on the same workload before changing production traffic.

Benchmark Routing lets users rank available models against up to three selected benchmarks with user-defined weighting. Requests are then sent to the highest-scoring model under that ranking.

Ramp also includes NVIDIA Switchyard routing, intended for agentic workflows in which the level of difficulty changes between steps. In that setup, routine steps can be handled by a lower-cost model, while more demanding tasks are escalated to a more advanced option.

Ramp gave one internal example from a coding agent run in which Switchyard handled 58 turns with one model and sent only the final five to another. Both prompt caches stayed warm during the process.

Cost focus

The new service arrives as companies seek closer control over AI spending, particularly when applications process high token volumes and face rate limits, variable latency and shifting provider pricing. Router shows users how each request was handled, including the model, provider, service tier, token usage, latency, cost and fallback attempts.

Ramp presented that visibility as a way for teams to understand why a request cost what it did and whether a cheaper or better-suited route was available. Routing decisions can include keeping a smaller model in place when a task does not justify escalation, or validating a less expensive alternative through shadow routing before any production change.

A customer cited in the launch said the product had simplified those choices. "It's just dead-simple. Between Flex tier and Switchyard this is free money with 0 effort, and it's saving me the headache of having to think about constantly switching models," said Josiah Parappally of Arcanist.

Another user highlighted the effect on operating costs. "Choosing the right model makes a meaningful difference to our AI spend. We run billions of tokens through Router, and have reduced our model costs by 92%," said Valentin de Matos of Delphi.

Broader market

Ramp is entering a growing area of the AI software market focused on orchestration rather than model development. Companies using large language models often need to balance quality, latency, reliability and cost across multiple vendors, particularly as newer models can outperform rivals in one task while lagging in another.

By positioning itself as model-neutral, Ramp is seeking to act as an intermediary between application developers and the model providers competing for usage. It does not build models itself and therefore says it has no reason to favour one over another.

That stance may appeal to businesses wary of locking applications to a single provider while the underlying market changes rapidly. Ramp says the aim is to let teams continuously optimise model selection without repeatedly rebuilding applications or running fresh evaluation projects from scratch.

Router now carries Ramp's own production AI traffic, and its routing methods were shaped by operating requirements including quality, latency, reliability and cost.