Ramp Router: your expense vendor now sells you inference

Ramp launched Router.com today: one API endpoint that sends each AI request to the lowest-cost model meeting your performance bar. Four days after Stripe agreed to buy OpenRouter, the expense-management company that watches your AI spend decided to sell you the AI. That is the story. Model routing stopped being infrastructure and became a customer-acquisition channel.

What actually happened

Ramp's launch post describes four routing strategies:

• Flex tier — drops to a provider's discounted service tier when latency is close enough to standard.
• Shadow models — runs a candidate model against real production traffic without touching what users see.
• Benchmark routing — weights up to three metrics and routes on the blend.
• NVIDIA Switchyard — keeps routine turns on a cheap model and escalates only the hard ones. Ramp's internal example: a coding agent ran 58 turns on the cheaper model and escalated the last five.

One API reaches OpenAI, Anthropic, and open models including DeepSeek, Kimi, GLM, and Qwen; the announcement says Gemini is coming. Ramp claims early customers cut inference costs about 40% on average. That is a vendor number about vendor customers, not an audited benchmark — treat it as the top of the range.

Routing is free through 2026 with $26 in model credits; you pay provider rates for tokens. 2027 pricing is unannounced. US only. Per TechCrunch, the default retains inputs and outputs for one year with PII stripped before product use.

Why model routing belongs in your stack, not your vendor's

The economics are real. Most teams pin one expensive model for every call because switching means a rewrite, and the cheap-model-plus-escalation pattern genuinely works. We build it into client systems.

The part to think about is where the seam lives. A router is the exact place where your prompts, your outputs, and your unit economics all pass through one throat. Handing that to a company whose other product already reads your spend is a real decision, not a checkbox.

Three questions before you point production at anyone's endpoint:

Is the interface portable? If your code calls an OpenAI-shaped API, you can swap the base URL when 2027 pricing lands. If you adopt proprietary routing verbs, you cannot. What happens to prompts? One year of retained inputs and outputs is a data-governance fact your contracts may already speak to. Change the default or accept it in writing. Who owns the eval? Routing decisions are only as good as the benchmark behind them. If the vendor grades the models and also bills the tokens, keep your own scorecard.

Sources: Ramp — Introducing Ramp Router, Ramp press release, TechCrunch.