# Ramp Router Ramp Router (router.com) is an LLM gateway launched on 2026-08-20 by Ramp, the US corporate-card and spend-management company. It exposes a single OpenAI-compatible endpoint in front of models from [[OpenAI]], [[Anthropic]], [[xAI]], DeepSeek, Moonshot, MiniMax, Z.ai, Qwen and NVIDIA, and routes each request to the cheapest approved model that still clears a configured quality bar. Ramp claims a 40% average cut in inference cost. It competes directly with [[OpenRouter]], but the interesting part is the business model rather than the routing: Router feeds Ramp's AI spend visibility product. The router is a wedge into corporate AI budgets, which is why routing itself is free through 2026. ## What is distinctive - **Routing strategies, not just fallbacks**: a Flex tier for discounted provider capacity, benchmark-weighted model ranking, prompt caching, and shadow models that mirror a slice of production traffic to candidate models to compare real cost and latency before switching. - **NVIDIA Switchyard**: an opt-in strategy for coding-agent traffic ([[Claude Code]], Codex and similar) that reads conversation signals — error severity in tool outputs, tool-call categories, test results, turn depth — and escalates to a frontier model only on complex or error-heavy turns. - **Two API surfaces**: `POST /v1/responses` (OpenAI Responses) and `POST /v1/messages` (Anthropic Messages). Chat Completions is not supported. ## The trade-off worth knowing Content recording is **enabled by default**: inputs, outputs, and tool calls are recorded and retained for one year, used for debugging, semantic analysis, and product improvement. Opting out is account-wide, propagates asynchronously, and does not delete what was already captured. That is the inverse of OpenRouter's logging-off default. Two smaller traps compound it. BYOK covers only OpenAI, [[Fireworks AI]] and xAI — Anthropic and Google Vertex traffic always runs on Ramp's shared credentials. And shared-key fallback is on by default, so a rate-limited or failing request on your own key silently moves onto Ramp's contract unless you disable it per provider. A cost router is a data router. The saving is real, and so is the fact that every prompt now has an extra custodian with a one-year default retention. ## References - Official website: https://router.com/ - Documentation: https://docs.router.com/ - FAQ (retention, hosting, BYOK): https://docs.router.com/resources/faq - Supported models: https://docs.router.com/supported-models - Launch announcement: https://ramp.com/blog/router-launch ## Related - [[OpenRouter]] - [[Large Language Models (LLMs)]] - [[OpenAI]] - [[Anthropic]] - [[xAI]] - [[Fireworks AI]] - [[Claude Code]]