Ramp launched its own AI model router.
By Joe Shelerud · September 2, 2026 · Curated by George's Blog
Ramp launched its own AI model router.
They ran it internally for three years and turned it into a product called Router. One API into OpenAI, Anthropic, DeepSeek, xAI, Nvidia and a few others. Route by cost, by benchmark, or by whoever has capacity that hour. Free through the end of 2026, inference is still on you.
The dashboard isn't why this matters.
Betting your whole stack on one provider looks riskier every month and the cost of getting out isn't just the API call. The cost is everything wrapped around it. Prompts tuned to one model's quirks, evals built on one provider's output, tool calling formats, etc..
A router is where you pay that cost once instead of every time.
The other half is that the models got good. Most of what we actually run isn't hard. Classification, extraction, pulling structure out of messy text, summarizing. A small model handles that fine now, and the price gap between it and the frontier model is often an order of magnitude.
Anyone here moved real production volume down to a cheaper model or use a router? Curious how that went.