Skip to content
NewsProduct launch

Ramp launches Router.com, a free AI model routing API, claiming 40% average inference cost savings

· by Pondero Newsdesk

The short version

Ramp opened its three-year-old internal AI routing infrastructure to the public on August 19, offering a free API that routes inference requests across 8 providers and claims a 40% average cost reduction.

Ramp launches Router.com, a free AI model routing API, claiming 40% average inference cost savings

Ramp, the corporate expense management company, opened its internal AI inference routing infrastructure to the public on August 19, three days after Stripe announced its $7 billion-plus acquisition of OpenRouter. The product, Router.com, is a single API endpoint that routes LLM requests across eight model providers and is free to use through the end of 2026. Per Ramp's launch post, the system has saved Ramp customers 40% on AI inference costs without reducing quality or uptime.

What

Router.com supports models from OpenAI, Anthropic, DeepSeek, Moonshot, Minimax, Nvidia, xAI, and Z.ai, per the Ramp announcement. Four routing strategies handle different trade-offs. The Flex tier strategy monitors a provider's discounted service tier in real time and automatically routes requests there when observed latency matches standard pricing, passing the discount through to the caller. Shadow models runs a candidate model against a sample of live traffic in parallel with the current production model so teams can compare cost and output quality before committing to any change. Benchmark routing selects the model that scores highest across up to three user-specified benchmarks. NVIDIA Switchyard keeps routine agent steps on a lighter model and escalates only the harder turns to a frontier one. In Ramp's own coding-agent logs, Switchyard handled 58 turns with Luna and sent only the final five to Sol, keeping both prompt caches warm throughout.

The service launched with a $26 credit and is currently US-only, per TechCrunch. Pricing for 2027 has not been announced. Router logs model inputs, outputs, and tool calls for one year by default; users can opt out, and Ramp says it strips personally identifiable information before using those logs to improve the product.

Why it matters

Ramp built Router on three years of its own production inference workloads, per the company, so the 40% savings figure comes from real fintech traffic rather than a vendor-arranged benchmark. Ramp also notes it builds no AI models and has no financial incentive to favor any provider, a position it says sets Router apart from routing services attached to a cloud or a lab. For teams with a growing monthly inference bill, Router is a free experiment through December 2026.

The Switchyard strategy has a direct cost-curve payoff for agent builders. A multi-turn agentic workflow that currently sends every step to a frontier model can stay on a cheaper model for most turns and escalate only when it must, cutting costs without changing output quality on the steps that need capability.

The main constraint is catalog depth. Per TechCrunch, OpenRouter currently offers many more model options than Router's eight-provider lineup.

What to watch next

Ramp has not announced 2027 pricing, so the cost calculus will shift once a routing fee is added. The next signal is whether Stripe uses the OpenRouter acquisition to deepen catalog access or announce tighter Stripe Billing integration. Either move would force any team evaluating Router.com to rerun its assumptions before committing.

Sources