Ramp launched Router.com, a new API-endpoint on August 19th, 2026. This is a platform for corporate spending that manages more than 200 billion dollars in annual purchases. Each AI request will be sent to the most cost-effective model meeting the developers’ quality standards. The service has already helped customers reduce their inference costs on average by 40%, according to the company’s announcement. Users pay the list price of tokens consumed and receive $26 credits for new users.
Ramp has released the tooling Ramp created for themselves three years ago. The company claims that by matching internal jobs to the correct model it has reduced its inference costs for each job around 30% while maintaining 99.9%+ of reliability. The Router site The current monthly production is more than 2,75 trillion tokens.
“AI is the fastest-growing line item at most companies, and the one they can least measure,” Rahul S. Sengottuvelu is Ramp’s Chief Technology Officer. “Router puts every token in one place and sends each request to the model that delivers the right performance at the right cost.”
Ramp tracks its own data to determine the timing. The Ramp AI IndexAccording to the company’s announcement, AI-related spending has grown by 20.7x in the last 20 years.
There are 27 models of the One Endpoint.
Gemini support is coming soon. Fireworks AI is serving open-weight models such as Nvidia’s Kimi, DeepSeek and GLM. Google, AWS and Together AI are also on their way. This is the model selector on Ramp’s Router page Currently there are 27 different models available, ranging in price from GPT-5.6 sol and Claude Opus 5 to GPT 5.4 nano and DeepSeek v4 Flash. Grok is now available on Amazon Bedrock, extending a day-old distribution campaign that saw Grok version 4.6 become generally available.
Routing engine performs more than 100 optimizations in model selection, caching and compression, timing, request handling, as well as automatic fallbacks when providers fail. Ramp can be configured to accept the default route or you can set your own priority for cost/performance per request. The models run on U.S. infrastructure with options for zero data retention.
It is possible to build a benchmark from production work
Ramp SWE Bench is the piece that carries all of the weight. This benchmark was created based upon real engineering production tasks at the company, rather than on public leaderboards. Router continually tests new models with that workload. It automatically folds in the best performers. In the published table, the router shows its spread: Claude Opus solves work on average for $1.84 whereas Qwen3.7 Plus reaches a solve rate comparable at $0.15 while GPT 5.4 Nano is able to handle cheaper tasks at $0.09.
As the model menu has expanded, inference economics is moving toward this type of per-request arbitration. The benchmarked serving costs for models that have comparable solve rate now differ by over an order of magnitude. Ramp’s table SWE Bench and other cheaper-inference play like Runware’s containers data centers reflect this spread.
The Terms and the Reports of Customers
On the product page, early users report that they have saved more than 40% on average. Delphi’s Valentin De Matos says that his company has run billions of tokens on Router, and model costs have been reduced by 92%. Katelyn Leses, Anthropic’s director of platform engineering said Claude will be available on Router as soon as it is released. SpaceXAI’s CEO said Grok’s addition expands the ways teams can integrate its models.
Ramp did not announce pricing after 2026. Router saves model inputs outputs metadata to help improve service. However, users can opt out of this by changing settings. At launch, the service will only be available to U.S.-based developers and teams. Enterprise features and other countries are expected soon. Ramp states that Ramp cards or company accounts are not required. Users of OpenAI SDK and Anthropic SDK can switch an existing integration by changing the base URL.

