Vercel adds beta routing rules to AI Gateway

Vercel has added beta routing rules to AI Gateway, introducing centralized controls for how model requests are handled across applications.

What changed

In a Vercel changelog post titled “Routing rules now available on AI Gateway,” the company says AI Gateway now supports gateway-level routing rules in beta. The feature is designed to let teams manage model usage from the gateway rather than by editing each application that sends requests.

According to Vercel, the initial rule types include rewrite rules and deny rules. Rewrite rules can transparently reroute requests from one model to another. Deny rules can block requests to specified models and return a 403 response.

That means a team could redirect traffic away from a model for cost, availability, or policy reasons, or prevent use of a model that has not been approved internally. Vercel describes the controls as operating at the gateway level, so applications can keep making requests while the routing behavior is changed centrally.

Why it matters for AI infrastructure

AI applications often depend on multiple model providers and model versions. As usage grows, teams need ways to manage provider changes, model deprecations, cost controls, and internal compliance requirements without shipping new application code for every routing decision.

The Agentic Digest, in its coverage of the update, characterized the feature as a way for teams to centrally control model use and reroute traffic without code changes. AI Primer similarly described Vercel’s routing rules as “firewall-style” model controls, with rewrite and deny behavior managed through Vercel tooling.

Those descriptions align with Vercel’s own changelog: the feature is not presented as a new model or inference engine, but as a policy and routing layer for AI Gateway traffic.

How teams may use it

The most direct use case is model substitution. If a team wants requests for one model to be served by another, a rewrite rule can handle that reroute at the gateway. This can help during migrations between model versions or when a team wants to test a replacement model without changing the calling application.

Deny rules address a different need: restriction. Vercel says blocked model requests return 403 responses. That gives platform teams a way to enforce model access policies, such as preventing use of experimental, expensive, deprecated, or unapproved models.

Megachangelog lists Vercel’s July 2 feature update as adding AI Gateway routing rules for model-level control without application code changes. That framing is consistent with the broader infrastructure trend around centralizing AI operations controls as more applications begin calling external models.

Availability and limits

Vercel’s changelog identifies routing rules for AI Gateway as a beta feature. The company’s announcement specifically names rewrite and deny rules as the supported rule types in this release.

Because the feature is in beta, teams evaluating it should check Vercel’s current documentation and CLI behavior before using it for production-critical policy enforcement. The sources provided do not specify general availability timing, pricing changes, or provider-specific limitations.

For now, the update gives Vercel AI Gateway users a more centralized way to steer and restrict model traffic, reducing the need to hard-code model routing decisions inside individual applications.