Export AI Gateway traces with Vercel Drains

Vercel AI Gateway now creates an OpenTelemetry trace for every request it processes. These traces can be exported through Vercel Drains to any endpoint that accepts OTLP over HTTP. The documentation lists native integrations for Braintrust, Dash0, Kubiks, Sentry, and Statsig. Access to this feature is limited to teams on the Pro and Enterprise plans.

To enable the export, an engineer configures an AI Gateway Trace Drain, selects the projects to monitor, and sets a sampling rate. Each trace captures the full request lifecycle: model and provider selection, any fallback or retry attempts, token usage and associated cost, time to first token, total request duration, and the final response status. Traces also include attribution fields such as project, deployment, API key, environment, and custom tags. The trace payload deliberately omits the prompt text and the model completion. Sampling controls allow the team to decide what fraction of traffic is sent to each drain.

AI Gateway traces are billed at $0.05 for each group of 1,000 traces successfully delivered to a drain. In addition, the standard Drains charge of $0.50 per gigabyte of transferred data applies. A request counts once per successful drain delivery, even if the gateway attempted multiple providers before succeeding. Failed delivery attempts do not incur a trace charge.

Delivered trace events can be viewed by count, and broken out by project, drain, or source. Engineers can set up a Trace Drain from the team’s Drains settings page or follow the AI Gateway documentation for detailed steps. The feature is unavailable to lower-tier teams; only Pro and Enterprise customers can use it. Because traces exclude prompt and completion content, they provide observability while addressing privacy concerns.