Rahul Sengottuvelu (@rahulgs) opened a closed beta for Ramp Router on July 20th, giving developers a single gateway for directing AI requests across an approved catalog of models while tracking the cost of each request.
Ramp built Router to manage more than 100 AI use cases inside its own products, Sengottuvelu wrote in the launch thread. He said the system cut Ramp's large language model costs by 30% while improving the speed and intelligence of its features. Ramp has not published the underlying cost data or evaluation results, so both performance claims remain company-reported.
The product exposes an OpenAI-compatible API, allowing developers to adopt it without rewriting each application around a new interface. Launch materials describe a managed model catalog, usage and spend reporting for individual requests, and tests that match requests to models based on semantic fit.
That routing layer addresses a problem Sengottuvelu described as a moving "price-intelligence-latency frontier." A model selected for a production feature can quickly become the wrong choice as providers release cheaper, faster or more capable alternatives. Keeping fixed model assignments leaves developers paying for intelligence a task does not require or accepting weaker output to protect margins.
Ramp Router moves that decision out of application code and into shared infrastructure. In a reply to former Cohere.io co-founder Yunyu Lin (@yunyu_l), Sengottuvelu said Codex should work with the service and Claude support is coming. Ramp has not disclosed Router's pricing, complete model catalog, beta capacity or the methodology behind its semantic-fit tests.
From tracking AI bills to controlling them
Router extends Ramp's push beyond corporate cards and expense software into AI infrastructure. Ramp released AI Spend Intelligence earlier in 2026 to combine provider usage, invoices, software subscriptions and card purchases. On July 16th, Ramp expanded those controls with weekly spend analysis, anomaly detection and recommendations for reducing waste.
The gateway gives Ramp a place to act on that information before an invoice arrives. If Router can connect a request's business purpose, model choice and final cost, Ramp can sell engineering teams the execution layer and finance teams the controls around it. That pairing is the strategic point of the launch: Ramp already monitors what businesses pay for AI, and Router puts Ramp directly in the path of the requests creating that spend.
The move also puts Ramp into an established market. OpenRouter offers a unified API for hundreds of models, automatic fallbacks and cost-aware selection. Vercel's AI Gateway provides access to hundreds of models with budgets, usage monitoring, load balancing and fallbacks. Vercel added gateway-level routing rules on July 2nd, allowing administrators to block models or redirect traffic without changing application code.
Ramp's opening is its existing position inside corporate finance. Router is being introduced as an extension of spend governance rather than a standalone model marketplace. The approved catalog and request-level accounting could appeal to companies where developers want rapid access to new models while finance and security teams require centralized controls.
Sengottuvelu turns an internal system outward
Router is Sengottuvelu's first external product launch since Ramp named him CTO on June 25th. He joined Ramp through its 2023 acquisition of Cohere.io, the customer-support automation startup he co-founded with Lin.
Before Cohere.io, Sengottuvelu worked in applied AI at Meta and on interfaces at Superhuman. He also created jsonformer, an open-source library for generating structured JSON from language models. Cohere.io had raised $3.5 million from Initialized Capital, Y Combinator and other investors before Ramp acquired its six-person team in June 2023.
Cohere.io built customer-service agents during the GPT-3 era, when production teams were already confronting the gap between impressive model demonstrations and predictable software behavior. Router applies the same operating instinct at a larger scale: model capability changes too quickly for every product group to maintain its own selection logic.
The 30% savings claim gives Ramp a clear beta pitch, though adoption will depend on whether Router can preserve output quality across real production workloads. Model routing can lower a bill simply by choosing cheaper models. The harder work is proving that those substitutions still complete each task at the required accuracy and latency. Ramp's own 100-plus use cases give Sengottuvelu a substantial internal test bed. The closed beta will test whether that system transfers cleanly to applications Ramp did not build.