Vercel AI Gateway Virtual Models let you swap the model behind your app without redeploying

Vercel has added Virtual Models to AI Gateway. A virtual model is a name you invent for a task, such as vmc/coding-agent. Your app or agent uses that name instead of a real model name, and you decide in AI Gateway which model sits behind it.

Today, switching models means editing the model name in your code and redeploying the app. Vercel says 125 new models from more than 10 labs reached AI Gateway in the last 90 days, about 10 a week, so that edit comes up often. With a virtual model, you change the setting in Vercel instead, and Vercel says your apps and agents pick up the new configuration within a few minutes, without code changes or redeploys.

What the virtual model holds

You create one from the command line or the team dashboard. Vercel’s example points a virtual model called coding-agent at anthropic/claude-opus-5.5 and turns on fast speed and Zero Data Retention:

vercel ai-gateway virtual-models create coding-agent \
  --model anthropic/claude-opus-5.5 \
  --config '{"speed":"fast","zeroDataRetention":true}'

Zero Data Retention means the provider is not supposed to keep your prompts after answering. Fallbacks, which tell the gateway which model to try if the first one fails, and tags for tracking usage also live on the virtual model. Vercel says these settings stay in place when you switch the underlying model. A tag set on the virtual model replaces any tag sent with the request.

In an app built with Vercel’s AI SDK, the virtual model ID goes where the model name used to go. For coding agents such as Claude Code and Codex, Vercel says to connect them with vercel ai-gateway setup, then pick the virtual model in the agent’s model picker.

One setting now changes every place the ID is used

The convenience has a catch. If three features share one virtual model ID, changing it changes all three at once, including the cost and the data-retention behavior. A switch to a cheaper model, or to one served by a provider that keeps prompts, takes effect everywhere with no deploy to review. Give each task its own ID so a change to your coding agent does not alter a customer-facing feature.

End of article