Claude API adds OpenTelemetry export in beta, with SDK and API traces

what changed

On 14 October 2026, Anthropic launched OpenTelemetry export in beta. You can now trace Messages API requests and Claude Managed Agents sessions in your own observability backend.

There are two parts. The Claude SDK emits spans in Python, TypeScript, C#, Go, Java, PHP and Ruby. For spans from the API itself, you add an OpenTelemetry destination in the Claude Console. You can set up either part, or both. With both, one request shows its SDK spans and API spans in a single trace.

The docs say the spans use the GenAI semantic conventions, so platforms that support those conventions can read them without extra setup.

why it matters

Most teams that run Claude in production already trace their own services. Until now, the model call was often a blind spot, or a hand-built wrapper. Standard spans let you see a slow or failed request next to the user action or feature that caused it.

This helps most with agents. A Managed Agents session is many steps, and a trace is the natural way to see where time and tokens went. It is also a source of real traffic for building evals.

It is a beta. Expect the details to change, and read the terms the docs point to before you send production data.

what we would do

1. Turn on SDK tracing in a staging service first and check the spans reach your backend.
2. Add the API destination in the Console and confirm SDK and API spans join into one trace.
3. Build one dashboard for latency, errors and token spend per feature.
4. Check what your traces contain before you send them to a shared tool, and agree who may see them.

If you want help wiring this into an agent system, see our build service.