Ramp launches its own AI model router, called Router

Ramp launched Router, an API‑based model routing service that lets users select and switch among several large language models. The offering is limited to the United States, free through the end of 2026 (users remain responsible for the underlying inference costs), and includes a $26 introductory credit. Ramp says it has operated an internal version of this router for its own AI workloads for about three years.

Router provides access to models from OpenAI, Anthropic, DeepSeek, Moonshot, Minimax, Nvidia, xAI, and Z.ai. It includes several routing strategies—for example, preferring a provider’s flexible usage tier or choosing a model based on up to three user‑specified benchmarks—and a dashboard that reports token spend, cost, latency, fallback attempts, and related metrics. Data retention is opt‑out; inputs, outputs, and tool calls are stored for one year after personally identifiable information is removed.

• Supported models: OpenAI GPT‑4o, Anthropic Claude 3, DeepSeek V2, Moonshot, Minimax, Nvidia Nemotron, xAI Grok, Z.ai
• Routing strategies: provider flex‑tier preference, benchmark‑based selection (up to three metrics), cost‑aware fallback, difficulty‑based escalation
• Data policy: opt‑out retention of inputs/outputs/tool calls for 12 months after PII removal
• Pricing: free access through 31 Dec 2026; $26 launch credit; users pay only for underlying model inference

Why this matters

By adding a model‑routing layer to its expense‑management platform, Ramp can capture a share of the growing AI inference market while offering customers a way to monitor and control token spend across multiple providers. The opt‑out data retention policy addresses privacy concerns, and free‑through‑2026 pricing lowers adoption barriers for cost‑sensitive enterprises.