Virtual models now available on AI Gateway

Vercel has launched Virtual models on its AI Gateway platform, introducing centralized configuration profiles that govern how external applications and autonomous coding agents interact with large language models. As detailed in Vercel Blog's report, a virtual model operates as a saved blueprint in AI Gateway that defines which foundational model and operational parameters incoming requests must use. Instead of embedding direct upstream provider references inside client codebases or local agent environments, software engineers point their systems to a custom identifier using the vmc namespace, such as vmc/coding-agent. When updates to models or parameters are needed, teams adjust the definition centrally within AI Gateway, eliminating the need to redeploy production services or reconfigure distributed developer tools.

This coverage examines the technical mechanics of Vercel's Virtual models, explaining how the architecture decouples client-side invocations from vendor-specific endpoints. The feature is targeted at software engineers, platform architects, and engineering managers running AI SDK applications or multi-agent coding workflows across development teams. By abstracting model endpoints behind stable slugs, organizations can standardise routing policies, enforce strict compliance measures like Zero Data Retention, and streamline cost allocation without disrupting day-to-day coding activities.

What shipped in Virtual models

Vercel released Virtual models as a native configuration layer across AI Gateway, making the capability compatible with every model type supported by the platform. The core mechanism allows developers to save an operational profile under a custom slug, which is then exposed as a dedicated model endpoint that accepts incoming completions and agent calls. Client integrations invoke this abstraction through the AI SDK or compatible API frameworks using the vmc prefix followed by the chosen identifier. This setup routes requests through AI Gateway, where preselected settings are injected and applied before the payload reaches the target foundation model.

The release incorporates direct support for developer-focused coding agents, allowing autonomous programming assistants to leverage centralized configurations. Teams can generate virtual model profiles either through the graphical dashboard on the Virtual Models page or programmatically via the Vercel CLI. By routing agent calls through virtual identifiers, platform operators can establish standard execution environments for development tasks. The update ensures that changes to underlying model weights, upstream providers, or operational toggles propagate immediately across all connected client interfaces without requiring manual intervention from developers.

What improved in Virtual models

Virtual models resolve the operational friction of updating model endpoints and system options across independent application repos and local agent installations. Prior to this update, modifying an agent or application endpoint required committing new provider identifiers, updating local configuration files, or executing full redeployments of production applications. With Virtual models, configurations reside within AI Gateway, allowing teams to swap foundational models, configure fallback models, and adjust operational parameters without touching application source code.

The architecture also expands parameter control for specialized tools that lack dedicated user interfaces for advanced model settings. Developers can preconfigure fast mode, prompt caching, and reasoning effort inside a virtual model profile, ensuring that agents benefit from these optimizations even if the agent software does not expose the relevant flags natively. This gives engineering organizations granular control over execution speed, cost containment, and context caching across every connected programming interface.

What you gain from Virtual models

Engineering organizations gain centralized routing, security governance, and compliance management across their entire development stack. Platform administrators can enforce compliance standards by applying requirements like Zero Data Retention directly within the virtual model configuration. Rather than trusting individual developers or local agents to configure data privacy settings correctly, AI Gateway guarantees that every request passing through the virtual model adheres to organizational compliance rules before contacting provider infrastructure.

Virtual models also introduce granular financial visibility through spend attribution tools. Organizations can attach custom reporting tags to any virtual model profile, which AI Gateway automatically binds to every downstream completion request using that identifier. This capability enables finance and platform teams to categorize API expenditures by project, repository, or operational team. By coupling financial reporting tags with fallback model definitions, teams can manage upstream downtime risks while tracking exact consumption patterns across autonomous agents and user-facing software.

How to get Virtual models

Developers can create a virtual model by navigating to the Virtual Models page in the AI Gateway dashboard and selecting Create Virtual Model, where they enter a slug, assign an underlying model, configure preferred parameters, and save the profile. Alternatively, teams can provision profiles via the Vercel CLI using dedicated management commands.

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

To integrate with coding agents, developers connect their environment to AI Gateway by running the setup command in their terminal.

vercel ai-gateway setup

Once the virtual model exists and local tooling is connected, developers update their applications by substituting the provider model ID with the newly created virtual slug. In AI SDK implementations or compatible API payloads, the model string is replaced with the custom slug format.

vmc/coding-agent

For coding agents, developers open the agent model picker or local configuration file and select the virtual model alias directly, ensuring all future queries route through AI Gateway with the configured parameters.

What to watch after Virtual models

Teams migrating to Virtual models should observe latency and throughput across their configured endpoints to ensure that gateway routing and fallback rules perform as expected under heavy agent workloads. When configuring fallback models, engineers must verify that downstream systems handle token limit variations and distinct feature support across different model families without throwing validation errors. Monitoring request tags within AI Gateway analytics will help confirm that custom reporting tags populate accurately across all connected projects.

Organizations should also review data retention flags across their deployed profiles to confirm that endpoints handling sensitive codebases consistently enforce Zero Data Retention. As new foundation models become available on AI Gateway, administrators can run canary evaluations by shifting a subset of agent traffic to updated slugs, validating agent performance and reasoning effort before rolling the identifier out across broader engineering divisions.