Creative finance-tech company Ramp entered the LLM gateway business this week with Router, an AI model routing service that gives developers one API to evaluate and switch between models from eight providers, with smart routing strategies that already cut Ramp’s own internal AI spend by 30%. The service is free through the end of 2026, and it lands right as Stripe is closing its $7.5 billion acquisition of OpenRouter — a direct signal that the model-routing layer is where the next wave of infrastructure value is forming.
OpenRouter launched Fusion a few months ago, and the multi-provider gateway space has been consolidating fast. Ramp’s entry is notable not because a finance company decided to build an AI product, but because it pairs routing with the token-cost telemetry Ramp already sells to thousands of businesses. The result is a router that is unusually honest about the thing most gateways bury: how much every routed call actually costs.
What changed
- Router is live now in the US, free through 2026 for the routing layer (you still pay provider inference costs) plus a $26 launch credit.
- One API routes to models from OpenAI, Anthropic, DeepSeek, Moonshot, Minimax, Nvidia, xAI, and Z.ai.
- Four routing modes, including flex-tier preference routing, benchmark-based routing (choose up to three benchmarks to sort models), cost-tier routing (send hard problems to expensive models, easy ones to cheap), and manual model testing without code changes.
- A dashboard exposes token spend, cost, latency, fallback attempts, and per-model usage.
- Ramp says Router cut its own AI costs by 30%; early customer reports put typical savings in the 40% range.
- Data retention note: Router records model inputs, outputs, and tool calls for one year by default, stripping PII before using them to improve the product. Opt-out data retention is a sharp difference from OpenRouter.
Router is a live gateway, and the difference between it and the routing you hand-roll is the routing strategy layer plus the cost telemetry on top.
How Router routes
At its core Router is the same pattern covered in my guide to multi-provider AI gateways and fallback routing: a single endpoint in front of many upstream providers, with a strategy deciding which upstream model handles each request. What Ramp layers on is the decision logic and the numbers.
The four routing strategies are worth understanding because they map to real engineering trade-offs:
| Strategy | What it does | Best when |
|---|---|---|
| Flex-tier routing | Prefers providers’ flex/batch usage tiers when available | Cost-sensitive workloads that tolerate slower responses |
| Benchmark routing | Ranks models by up to three user-selected benchmarks and routes to the top scorer | You care about a specific eval, not a single vendor default |
| Cost-tier routing | Routes hard/complex problems to expensive frontier models, easy ones to budget models | Mixed workloads where quality threshold matters |
| Manual selection | Test a specific model without rebuilding your app | Evaluation, A/B, migration planning |
The key insight: benchmark routing ties routing decisions to actual measured quality rather than a static “always use the frontier model” default. If your eval says a smaller model matches the frontier on your task, Router can route there and save you the margin every call.
Cost math that matters
Ramp’s 30% internal figure and the 40% customer-reported figure both come from the same mechanism: stop defaulting every request to the most expensive model. The savings stack up like this:
- Easy queries routed to budget models: if 60% of your traffic is classification, extraction, or short generation, sending it to a mid-tier model instead of frontier takes most of the cost out.
- Benchmark-based selection: when a mid-tier model scores within a point or two of the frontier on your eval, the routing cost is roughly a third to a half of frontier pricing.
- Flex-tier usage: batch and flex tiers charge meaningfully less per token, accepting latency in exchange.
Compare that to my write-up on the DeepSeek API price increase — when a single provider changes pricing, a router that can shift traffic instantly is a hedge, not just a convenience.
Decision framework
- When to use: you already work with multiple providers, your workload quality ceilings vary by task, and you want per-call cost visibility without building the routing and telemetry yourself.
- When not to use: you have strict data residency requirements (US-only availability, one-year retention), you need a niche model Ramp doesn’t carry, or you’re already deep in OpenRouter with complex existing routing.
- Trade-off: the routing convenience is real, but the gateways differ on model breadth, retention, and independence. OpenRouter carries far more model options today; Ramp’s edge is cost telemetry and the free-until-2027 runway.
- Recommendation: for teams that already run a multi-model stack and want cost control with minimal plumbing, Router is worth a trial now while it’s free. Keep an eye on the retention policy against your own compliance rules.
- Final takeaway: the model-routing layer is becoming a first-class infrastructure product — Stripe and Ramp betting $7.5B+ is the tell — and Router is a credible, cost-obsessed entry.
Implementation notes
- The Stripe/OpenRouter deal (Stripe acquires OpenRouter for $7.5B) shows the routing layer is centralizing; evaluate vendor lock-in risk before deep integration.
- If you route via Ramp’s dashboard today, export per-model cost and latency before building automations on top — the free pricing ends in 2027 and per-model pricing may shift.
- For teams already on my Laravel multi-provider AI routing pattern, treat Router as a managed alternative to your fallback chain rather than a replacement for your abstraction layer.
- Model breadth changes fast in this space; re-check the upstream provider list before committing a workload to Router.
Related reading
- OpenRouter Fusion: Match Fable 5 on Deep Research at Half the Cost — the fusion approach that turns many models into one answer, the router’s opposite bet
- Multi-Provider AI Gateways and Fallback Routing — the gateway architecture Router wraps in a managed API
- Laravel Multi-Provider AI Routing — how to build your own routing layer if you’d rather not depend on a third party
- DeepSeek’s Significant API Price Increase — why router flexibility matters when a single provider reprices
Sources
- Ramp blog — Introducing Ramp Router: Change AI Models Without Rebuilding — official announcement, strategies, pricing, 30% claim
- TechCrunch — Ramp launches its own AI model router, called Router — launch coverage, $26 credit, US-only, provider list
- TechCrunch — Stripe didn’t really buy OpenRouter because of the ‘singularity’ — $7.5B deal context and gateway-consolidation reading
- TechCrunch — Ramp raises $750M at $44B valuation — Ramp’s valuation context behind the AI push
About the author
Charles Jasthyn De La Cueva is a full-stack developer and the founder of Open TechStack. He writes about AI engineering, developer tools, and practical model evaluation — grounded in real workflows, not press releases.
