Vercel Adds Confidence-Based Decision Fallbacks to AI Gateway.

Vercel’s confidence-based decision fallbacks let AI Gateway send a decision request to a second model when the first model’s answer falls below a confidence condition.

According to Vercel’s changelog, the beta feature works with decision requests made through AI Gateway. Builders set the fallback model and its conditions inside providerOptions.gateway.models.

The primary model still answers first. If its response trips the configured condition, AI Gateway runs the decision again with the fallback model. Both stages are billed when an escalation happens.

Conditions cover Choice and Score questions. Boolean questions use a probability range instead. Multiple conditions can be combined, allowing the fallback to depend on more than one signal.

Vercel’s example routes a customer report saying they were charged twice and the app will not load. The decision model chooses between billing and technical teams. A confidence condition can escalate that classification to another model when the first answer is uncertain.

This sits alongside AI Gateway’s existing error fallbacks. Plain model names already listed in models continue to catch outright errors, while conditional objects handle low-confidence answers. Requests without a conditional object keep their existing behaviour.

The release follows Vercel’s other additions to the gateway, including Browserbase Search and Fetch for AI Gateway.

The fallback is added as a conditional object in the gateway’s model list. That object names the second model and defines when it should run. Existing model entries can stay in place for error handling.

Because escalation triggers a second decision, it adds another billed model call to that request. Requests that do not meet the condition finish with the primary model’s answer.

What are Vercel confidence-based decision fallbacks?

They are an AI Gateway beta feature that sends a decision request to a fallback model when the primary model’s answer meets a configured low-confidence condition. The fallback model then makes a second decision.

Which question types support confidence-based decision fallbacks?

The conditions support Choice and Score questions. Boolean questions use a probability range, and multiple conditions can be combined.

How much do confidence-based decision fallbacks cost?

When a fallback runs, both the primary and fallback decision stages are billed. The update does not provide plan availability or exact pricing details.