Claude API adds OpenTelemetry trace export in beta, off by default

Anthropic said on 14 October 2026, in its Claude Platform release notes, that it has launched OpenTelemetry export in beta. It lets a team trace Messages API requests and Claude Managed Agents sessions in the observability tool it already runs. OpenTelemetry is a vendor-neutral standard for traces: a trace is the record of one request, and a span is one step in it, such as a model request or a tool use.

What you get

Anthropic's documentation describes two sources of spans that can be used together or separately. The Claude SDK emits spans from your own code, and the Claude API emits spans from Anthropic's side. The docs say the API spans cover each model request and each tool the API runs, and that session turns are traced for Managed Agents. Both use the GenAI semantic conventions, so platforms that understand those conventions can read the spans without extra setup.

The documentation lists these partner destinations: Arize AI, Braintrust, Databricks, Datadog, Grafana Cloud, Honeycomb, Langfuse, LangSmith, New Relic, PostHog, Sentry and Splunk. Any platform that accepts OTLP, the OpenTelemetry wire protocol, can also receive them.

What leaves your account

By default each destination is set to metadata only: identifiers, timing, status, token counts and the names and descriptions you give agents, tools and MCP servers. Full content adds model prompts and responses. The docs say each span holds up to 0.75 MB of content by default, enough for a 200,000-token window of English text, and that a full 1-million-token context needs about 4 MB. Longer content is truncated and flagged with an attribute.

The docs state that once spans are delivered, the data sits with the platform you chose, outside Anthropic, under that platform's access and retention settings. They also say destinations export only metadata for organisations with a HIPAA, FERPA or K-12 data-handling restriction with Anthropic, whatever the content mode. With zero data retention, Full content is still allowed, and for Messages API requests the API holds the exported content in memory only until it is delivered; Managed Agents is not eligible for zero data retention.

How to switch it on

The API side is configured per organisation. Per the docs, an admin adds a destination in the Claude Console (or through the Admin API) with an OTLP/HTTP endpoint that accepts protobuf over HTTPS on port 443, any request headers for credentials, a content mode and the workspaces it covers: all, selected, or all except some. A Test button sends an anthropic.test span, and the Delivery column shows Healthy or Failing. SDK tracing needs no admin role, only a tracer provider registered in your application.

To see API spans nested inside your own traces, the docs say requests carry a W3C traceparent header, which the SDK writes for you; if the sampled flag is off, the API emits no spans for that request.

What the documentation leaves open

• Claude Platform on AWS: the docs say it does not yet support the traceparent header, so API spans there start their own traces.
• Cost: the pages I read do not mention a charge for exporting spans, and do not say how long the beta will last or what changes at general availability.
• Evidence: Anthropic reports no figures on overhead or on problems found with the traces. This is a capability announcement, not a measured result.

For a team running agents in production, the practical question is whether the destination's retention rules fit what Full content would put in it. Metadata only gives token counts and latency per request without storing any conversation text.