Introducing Virtual Models on AI Gateway

Vercel has launched Virtual Models on AI Gateway, enabling teams to abstract model configurations behind a single virtual model ID. This allows developers to update models, set fallbacks, and enforce policies like Zero Data Retention centrally. Changes to the virtual model apply automatically across all apps and agents using its ID, eliminating the need for redeployments or code modifications. The system simplifies model management by decoupling configuration from application logic.

In the past 90 days, AI Gateway added 125 new models from over 10 labs, averaging about 10 new models per week. AI Gateway already supports one-line model switching via provider IDs, but Virtual Models extend this by moving all configuration to the gateway. Teams can assign a virtual model ID to tasks like coding agents or support ticket triage, then adjust models and settings in AI Gateway without touching application code.

Virtual Models support purpose-built configurations for different tasks, such as cost-sensitive operations like support triage or high-reasoning tasks like legal reviews. Settings like fallbacks, Zero Data Retention, and tags are tied to the virtual model, not individual requests. Provider-specific settings, such as those for Anthropic or OpenAI, are enforced only when the virtual model points to that provider, ensuring consistency across tasks.

The Virtual Models page provides visibility into requests, spend, and success rates for each task. Teams can create virtual models via the Vercel CLI or dashboard, with role-based access controls restricting who can modify configurations. AI Gateway enforces centrally defined settings, such as Zero Data Retention or provider restrictions, even if requests attempt to override them. Virtual Models work with all model types on AI Gateway, including image, video, speech, and embedding models.