Export AI Gateway traces with Vercel Drains

AI Gateway now produces an OpenTelemetry trace for every request. Pro and Enterprise teams can export those traces through Vercel Drains to any OTLP/HTTP-compatible endpoint. Native integrations cover Braintrust, Dash0, Kubiks, Sentry, and Statsig, so teams can land the same request telemetry in the observability or evaluation stack they already use.

Each trace covers the full request lifecycle. That includes model and provider routing, fallback and retry attempts, token usage and cost, time to first token, and request duration. Because the format is OpenTelemetry and the transport is OTLP over HTTP, the drain path is not locked to a single vendor UI; any endpoint that speaks OTLP/HTTP can consume the stream via Drains.

For engineers building on multi-model routes, the hard part is often not calling a model but explaining what happened when latency spikes, cost jumps, or a fallback fires. Per-request traces that surface routing choices, retries, token spend, and TTFT give a single record to debug production AI traffic without stitching logs from the app, the gateway, and the provider by hand.

The feature sits at the junction of two markets: AI gateways that abstract providers, and observability platforms that already understand OpenTelemetry. Native hooks for Braintrust, Dash0, Kubiks, Sentry, and Statsig reduce integration work for teams already on those tools, while the generic OTLP/HTTP path keeps the export path open for other collectors and backends.

If you run AI Gateway on Pro or Enterprise, the practical next step is to wire a Drain to your existing OTLP/HTTP endpoint or one of the named integrations and confirm that routing, fallbacks, token cost, and TTFT appear as expected on live traffic. Watch whether your tooling treats these spans as first-class AI traces (cost and model fields included) or only as generic HTTP spans, and adjust dashboards and alerts around fallback rate, TTFT, and cost per request rather than latency alone.