Ramp launches Router to route enterprise AI workloads across models

Introduction

For companies deploying language models, choosing a model is only part of the operational challenge. Teams also need to balance quality, latency, reliability, and the cost of every inference request. Ramp, the corporate expense management company, has launched Router, an AI model routing service designed to give users one API for accessing and switching between several large language models.

Key points

• Multiple providers through one interface: Router currently offers models from OpenAI, Anthropic, DeepSeek, Moonshot, Minimax, Nvidia, xAI, and Z.ai.
• Configurable routing strategies: Customers can favor providers’ flexible usage tiers or let Router select a model using up to three user-defined benchmarks.
• Different models for different difficulty levels: Users can send only harder problems to more expensive models and test alternatives without changing their integration.
• Operational visibility: A dashboard reports token usage, cost, latency, fallback attempts, and other call-level details.
• Early commercial terms: The service is available only in the U.S. for now. Router fees are waived for the remainder of 2026, while customers still pay model inference costs; pricing for the following year has not been disclosed.

Ramp says it has used the routing system internally for roughly three years before releasing it as a product. Its basic proposition resembles OpenRouter, although Router currently supports fewer model options. For developers and enterprise teams, a common API can reduce the work of integrating separate providers and make model testing, replacement, and fallback behavior easier to manage.

Why it matters

The significance of Router goes beyond forwarding a request to a different model. As providers introduce increasingly varied price structures and performance profiles, companies need a repeatable way to decide which model should handle each task. Routing difficult queries to premium systems while sending routine work to cheaper alternatives could improve the balance between quality and spending. Benchmark-based selection also gives teams a framework for making model choices less dependent on individual developer judgment.

The launch fits closely with Ramp’s existing products for monitoring AI token usage and managing token spending. Router could give Ramp a way to sell current customers an integrated path from model access to cost visibility. It may also create relationships with AI labs and inference providers, potentially giving the company a foothold in a market adjacent to corporate expense management.

Data handling will be a significant consideration for enterprise buyers. Router records model inputs, outputs, and tool calls for one year by default, while providing an opt-out mechanism. Ramp says it removes personally identifiable information before using that content to improve the product. Organizations working with sensitive data will still need to assess the retention period, the effectiveness of the de-identification process, and the practical scope of the opt-out controls.

Router illustrates how model gateways are moving from a developer convenience toward a layer of enterprise AI operations. Its long-term prospects will depend on model coverage, pricing after the introductory period, data controls, and whether its routing strategies deliver measurable improvements in cost and reliability.

Source: TechCrunch AI