With new models arriving weekly, moving model choice out of code lets you change models, fallbacks and data-retention policy for agents without a release cycle.
Key takeaways · AI-distilled
Vercel counts 125 new models from more than 10 labs added to AI Gateway in 90 days, about 10 a week. Switching was already a one-line change, but that line lived in code, so each switch usually still required a redeploy.
When a virtual model is switched, fallbacks, Zero Data Retention and tags carry over, but provider options are per provider. Vercel notes an Anthropic-specific setting stops applying once the ID points at an OpenAI model, so those options need review on every provider switch.
Vercel suggests pointing a second virtual model, such as a candidate ID, at a new model and running evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition → before switching the production ID, then pointing the original back if quality drops. Callers pick up an edited virtual model within a few minutes.
Settings on a virtual model override the request: callers cannot change its provider order, fallbacks or service tier, and Zero Data Retention or no-training requirements hold even if a request sets them to false. Its tags replace request tags so spend lands on the task.
Per Vercel, virtual models cover image, video, speech, embeddingA list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.Full definition → and realtime models as well as text, and work with the AI SDK and the OpenAI- and Anthropic-compatible APIs. Coding agents like Claude Code and Codex can select them after running the gateway setup command.
Terms in this piece · Glossary
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
