Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarVercel Blog→ original

Vercel Lets You Swap LLMs in Production Without Redeploying: 125 New Models in 90 Days

Vercel has added Virtual Models to AI Gateway. Applications call a task ID such as vmc/coding-agent rather than a specific LLM. The model behind that ID, its fallback providers and its data retention policy are set in the gateway's configuration. Over the past 90 days, 125 new models appeared in the gateway. Previously, every model change meant a code edit and a redeploy. Now all applications pick up the new configuration within a few minutes.

AI-processed from Vercel Blog; edited by Hamidun News
Vercel Lets You Swap LLMs in Production Without Redeploying: 125 New Models in 90 Days
Source: Vercel Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceVercel Blog ↗
Editor-in-Chief
Key metric #1Virtual model settings take precedence over request settings: code cannot override them
Key metric #2Zero Data Retention applies even when a request explicitly turns it off
Key metric #3To test, create a second ID such as vmc/coding-agent-candidate and run evals against it

New models are now arriving faster than teams can adopt them. According to Vercel, 125 models from more than ten labs were added to AI Gateway over the past 90 days, roughly ten per week. Switching to any of them in the gateway takes one line, but that line lives in your code. As a result, every model change usually ends in an edit, a review and a redeploy. Vercel proposes removing that line from the code with Virtual Models.

What a virtual model is

A virtual model is a task ID that you use wherever a model name string used to go. For example, you create vmc/coding-agent and specify once in AI Gateway what sits behind it: which model, what speed, and whether Zero Data Retention (a ban on the provider storing requests) is required. You can create such an ID in the team dashboard or through the Vercel CLI.

After that, only this ID remains in the code. In the example from the announcement, a streamText call from the AI SDK receives model: 'vmc/coding-agent', and the code knows nothing else about the specific provider. When a better model becomes available, including one from a different provider, you change the setting with the edit command, and all applications and agents calling this ID pick it up within a few minutes. Nothing needs to be deployed.

This also works for ready-made coding agents such as Claude Code and Codex. They are connected with the command vercel ai-gateway setup, and the virtual model is selected from the tool's own model list.

A dedicated model for each task

Tasks differ widely in load. Support ticket triage may run hundreds of times an hour, where cost matters most. Legal contract review runs several times a day but requires strong reasoning. Such decisions used to be scattered across repositories. Now each task gets its own virtual ID, and its configuration can be changed independently of the others.

When the model changes, some settings stay in place: fallbacks, Zero Data Retention and tags. Provider-specific options, however, apply only to that provider. If you had a setting specific to Anthropic and redirect the ID to an OpenAI model, it will stop taking effect. Vercel explicitly advises reviewing provider options whenever the provider changes.

For production, a safe scheme is proposed. Create a second ID, for example vmc/coding-agent-candidate, point it at the new model and run your evals (quality check suites) against it. If quality drops after switching the main ID, it can be pointed back to the previous model.

Control for platform and security teams

Any team member can call a virtual model, but only those with the appropriate role can change its settings. This gives platform and security teams a central point of control. They decide which models are available and with what parameters, while still quickly giving developers access to new releases.

An important detail: the virtual model's settings take precedence over the request's settings. Calling code cannot change the provider order, the fallbacks or the service tier. If fallback models are defined for vmc/storefront-assistant, every application will use the same failover logic, even when the request specifies other ones. The same applies to privacy. If a virtual model requires Zero Data Retention or providers that do not train on request data, this holds even when the request explicitly sets zeroDataRetention or disallowPromptTraining to false.

Tags are also overridden. The virtual model's tags replace those in the request, so costs always attribute to the task regardless of how the application labels its calls. The Virtual Models page shows requests, spend and success rate for each task.

What this means in practice

Virtual models work with all model types in AI Gateway: text, image, video, speech, embeddings and realtime. The AI SDK and OpenAI- and Anthropic-compatible APIs are supported. According to Vercel, the gateway passes through provider prices without markup and charges no platform fee for inference, including when bring-your-own-key (BYOK) is used.

The limitation is worth noting upfront: you tie yourself to AI Gateway as a control layer, and changing configuration without a deploy means a configuration mistake reaches production immediately. That is why the candidate-and-evals scheme here is not a formality. For teams with several agents and dozens of tasks, this is a practical way to stop running redeploys for every new model while keeping control over privacy and costs.

Story

Key Questions Answered

What a virtual model is?

A virtual model is a task ID that you use wherever a model name string used to go. For example, you create vmc/coding-agent and specify once in AI Gateway what sits behind it: which model, what speed, and whether Zero Data Retention (a ban on the provider storing requests) is required. You can create such an ID in the team dashboard or through the Vercel CLI. After that, only this ID remains in the code.

What are the key details on A dedicated model for each task?

Tasks differ widely in load. Support ticket triage may run hundreds of times an hour, where cost matters most. Legal contract review runs several times a day but requires strong reasoning. Such decisions used to be scattered across repositories. Now each task gets its own virtual ID, and its configuration can be changed independently of the others. When the model changes, some settings stay in place: fallbacks, Zero Data Retention and tags. Provider-specific options, however, apply only to that provider.

What are the key details on Control for platform and security teams?

Any team member can call a virtual model, but only those with the appropriate role can change its settings. This gives platform and security teams a central point of control. They decide which models are available and with what parameters, while still quickly giving developers access to new releases. An important detail: the virtual model's settings take precedence over the request's settings. Calling code cannot change the provider order, the fallbacks or the service tier.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…