--- title: "Ramp Launches Router.com to Cut AI Bills by 40%" date: 2026-08-20 author: "Barry Elad" featured_image: "https://sqmagazine.co.uk/wp-content/uploads/2026/08/ramp-launches-router-com-to-cut-ai-bills.jpg" categories: - name: "Artificial Intelligence" url: "/artificial-intelligence.md" tags: - name: "News" url: "/tag/news.md" --- # Ramp Launches Router.com to Cut AI Bills by 40% Ramp launched Router.com on Wednesday evening, a single API endpoint that sends each request to the cheapest AI model clearing a developer’s quality bar. Customers already using it cut inference costs 40% on average, the company said. ## The Big Picture - Ramp opened Router.com to the public this week, giving developers one API that reaches models from OpenAI, Anthropic, and SpaceXAI. - Router picks the lowest-cost model that meets the performance level a developer sets, then fails over to another provider automatically. - Ramp says customers running on Router cut inference bills by 40% on average, a figure the company measured itself. - Router records model inputs, outputs, and tool calls for one year by default, and developers have to opt out. - Routing is free through the rest of 2026 and ships with $26 in credits, with no price announced for 2027. ## Router picks the cheapest model that clears the bar Developers connect to [Router](https://router.com/) through one API and reach models from **OpenAI**, **Anthropic**, **SpaceXAI** (SpaceX’s AI division, formerly xAI), **DeepSeek**, **Moonshot**, **Minimax**, **Nvidia**, and **Z.ai**, with Google’s Gemini listed as coming soon. Ramp says the service runs more than **100 optimizations** across model selection, caching, compression, timing, and request handling, and switches providers automatically when one fails. > Ramp has launched its own AI model routing service, dubbed Router, that lets users and companies use and switch between various large language models via an API. > > — TechCrunch (@TechCrunch) [August 20, 2026](https://x.com/TechCrunch/status/2090481764091265043?ref_src=twsrc%5Etfw) Teams can accept Ramp’s default strategy or configure their own: one option favors providers’ flex usage tiers, another routes on up to three benchmarks the customer names, and a third sends only hard problems to expensive models. A dashboard tracks token spend, cost, latency, and fallback attempts. Router runs in the United States only. **Rahul Sengottuvelu**, Ramp’s chief technology officer, tied the launch to the spend products Ramp already sells. Sengottuvelu said: “ AI is the fastest-growing line item at most companies, and the one they can least measure. Router puts every token in one place and sends each request to the model that delivers the right performance at the right cost. Rahul SengottuveluChief Technology Officer – Ramp ## Ramp benchmarks models on its own engineering work Ramp built the router for internal use three years ago and says it cut the company’s own inference costs roughly **30%** for the same output while holding reliability above 99.9% across production traffic. Both that figure and the 40% customer average come from Ramp itself; no outside auditor has checked either. Model selection runs against [Ramp SWE-Bench](https://labs.ramp.com/swebench), a benchmark the company assembled from its own production engineering tasks, and strategies that score well there get folded into Router’s defaults automatically. The boundary is worth naming: it measures which model handles Ramp-style engineering work, a narrower question than which model suits a legal summarizer or a support bot. ## Router logs prompts for a year unless you turn it off Router records model inputs, outputs, and tool calls for a year under an opt-out policy, so the logging runs until a developer switches it off. Ramp says it removes “**personally identifiable information before using that content to improve the product.**“ That default matters most for teams routing production traffic carrying customer records, contracts, or source code. Anyone who has already pointed an application at Router should open the dashboard and check the retention setting, confirm whether live prompts carry regulated data, and rotate any keys pasted into test requests. Switching retention off helps reduce exposure, though it does nothing about logs already written. ## Payments companies are building toll booths on inference Ramp’s launch landed a day after [Stripe’s purchase of OpenRouter](https://sqmagazine.co.uk/stripe-statistics/) drew coverage, putting two payments companies in the same business inside 48 hours. Both already sit between companies and their vendors, and inference has become a vendor bill large enough to meter: Ramp’s own AI Index puts corporate **AI spending at 20.7 times its June 2025 level**. Ramp raised $750 million at a $44 billion valuation in June, and Router hands it a doorway into engineering teams that never touch an expense report. Router’s math depends on a wide price gap between frontier models and open-weight ones such as [DeepSeek](https://sqmagazine.co.uk/deepseek-ai-statistics/) and Z.ai’s GLM, both of which it carries. Several questions stay open: - **What routing costs once the free period ends, which Ramp has not disclosed?** - **How the 40% average was calculated, and against which baseline?** - **Whether Ramp SWE-Bench scores will be published per model?** - **When Gemini and the named serving partners actually arrive?** - **Whether Router opens outside the United States?** ## SQ Magazine’s Takeaway **Router** is a cost-control product wearing infrastructure clothes. The routing layer is familiar, since OpenRouter has offered a comparable endpoint for longer and carries a wider model list; what Ramp adds is a wire between the routing decision and the finance team’s view of who spent what. For a company already running Ramp cards and bill pay, that wire is the pitch, and the 40% claim is the thing to verify on its own traffic. Pricing is what to watch. Free routing ends with 2026, Ramp has put no number on next year, and teams that build against Router in the meantime take on a switching cost at an unquoted rate. Engineering leads evaluating it can keep their own provider keys live, log a cost baseline before routing anything, and re-check retention after each dashboard change. No independent verification of the 40% figure exists yet, and buyers should read it as a vendor claim until one does. Definition of AI Inference. Link to full glossary entry follows the description.**AI Inference**AI inference is the execution phase where a trained AI model applies what it learned to new, unseen data and produces an output such as a prediction. [Read more](https://sqmagazine.co.uk/glossary/ai-inference/)