Vercel launches Virtual Models on AI Gateway for no-redeploy model switching

Vercel has launched Virtual Models on AI Gateway, an abstraction that decouples the model string in application code from the model and configuration that actually serves requests. According to Vercel's launch post, a virtual model is a saved configuration paired with a custom ID such as vmc/coding-agent. Applications and agents call that ID, and teams can swap the underlying model, adjust fallbacks and enforce retention rules inside AI Gateway — every caller picks up the change within a few minutes, with no code change or deployment.

The launch targets a specific pain point: model churn. Vercel says 125 new models from more than 10 labs arrived on AI Gateway in the past 90 days, roughly ten per week. The gateway already made switching a matter of editing a single line of code, but because that line lives in the codebase, a switch still usually meant a redeploy. Virtual Models move that decision out of the application entirely.

How teams use them

Virtual models are created from the team dashboard or with the Vercel CLI, where a command names the slug, the underlying model and a configuration block covering options such as fast mode and Zero Data Retention. The ID then replaces the provider model string in AI SDK requests and in the OpenAI- and Anthropic-compatible APIs. Coding agents such as Claude Code and Codex connect through the Vercel CLI setup flow and pick a virtual model from their model picker.

Vercel's changelog entry highlights a second benefit: exposing options that a given agent harness does not surface. Reasoning effort, fast mode and prompt caching can be set on the virtual model even when the calling agent offers no way to configure them.

Each task can have a virtual model matched to its workload — a high-volume, cost-sensitive job like support ticket triage can point at a cheap model, while a low-volume contract review task can use a heavyweight reasoning model. Settings such as fallbacks, Zero Data Retention and tags carry over when the model behind a virtual model changes, though Vercel cautions that provider-specific options do not: an Anthropic-only setting stops applying once the virtual model points at an OpenAI model, so provider options should be reviewed on every switch.

For production traffic, Vercel recommends a candidate pattern: point a second virtual model at the new model, run evaluations against it, and revert the original to the previous model if quality drops.

Central governance and spend tracking

Virtual models also act as a governance control. Team roles determine who may edit one, letting platform and security teams test and approve models centrally while everyone else consumes them by ID. Settings on the virtual model take precedence over the same settings in a request, so callers cannot override the centrally chosen provider order, fallback chain or service tier. Compliance works the same way: a virtual model that requires Zero Data Retention, or providers that do not train on request data, keeps those requirements even when a request tries to disable them.

Tags attached to a virtual model replace any tags sent with the request, so spend is attributed to the task regardless of how individual applications label their calls. The Virtual Models page reports requests, spend and success rate per task.

Why it matters

Model selection has become an operational decision as much as a development one. With new models shipping weekly, teams that hardcode model strings face a treadmill of edits, redeploys and reviews just to keep up — or to react when a provider degrades. By turning the model identifier into a centrally managed resource, Vercel shifts routing, failover and compliance closer to infrastructure configuration than application code, giving platform teams a choke point for governance and cost while application teams get a stable ID. The trade-off is added indirection: runtime behavior now depends on configuration outside the codebase, which makes the evaluation and rollback practices Vercel recommends essential. Virtual models work across every model type on the gateway — image, video, speech, embedding and realtime — and Vercel applies no markup or platform fee on inference, including bring-your-own-key requests.