Ramp, the corporate expense platform, has released Router. It is the model-routing service the company says it has been using internally for the past three years.
Router sends each request to a different large language model through one API. It covers OpenAI, Anthropic, DeepSeek, Moonshot, Minimax, Nvidia, xAI and Z.ai.
Four ways to route
You can route by preference, flexing between usage tiers. You can let it pick automatically using benchmarks, choosing up to three that matter to you. You can optimise for cost, so only the hard problems reach the expensive models. Or you can use it simply to test models without rewriting your code each time.
A dashboard reports token spend, cost, latency and fallback attempts.
The catch
It is free until the end of 2026, with a $26 launch credit. You still pay for inference separately, and Ramp has not said what it will charge afterwards. It is available in the United States only.
One detail worth reading before you connect production traffic: inputs, outputs and tool calls are retained for a year by default. Ramp says personally identifiable information is stripped before that data is used to improve the product.
Routing is where a lot of AI budget quietly goes. Paying frontier prices for work a small model handles adds up fast across millions of calls.
Source: TechCrunch AI
