Anthropic Adds OpenTelemetry Tracing to Claude Platform in Public Beta

Anthropic's Claude Platform OpenTelemetry integration is in public beta for Messages API requests, Claude Managed Agents sessions, and Claude SDKs. You can export traces to an observability platform that accepts OTLP, with metadata-only as the default content mode and a separate full-content option that can include prompts and responses. (Anthropic launch article; Claude Platform OpenTelemetry docs)

The feature gives developers a way to inspect model calls, tool use, and agent activity in their observability stack. The data mode matters: metadata-only is the default, while full-content tracing can expose prompt and response text to the configured destination. (Claude Platform OpenTelemetry docs)

Key Takeaways

• OpenTelemetry tracing for Claude Platform is in public beta. Anthropic lists the Messages API, Claude Managed Agents, and Claude SDKs. (Anthropic launch article)
• Traces can go to an OTLP-compatible observability tool. The launch article describes standard OpenTelemetry traces, while the docs explain API and SDK tracing setup. (Anthropic launch article; Claude Platform OpenTelemetry docs)
• Metadata-only is the default. Full-content mode can also export prompts and responses, so choose the mode deliberately. (Claude Platform OpenTelemetry docs)
• API tracing and SDK tracing cover different paths. Add an API destination in the Claude Console or configure a tracer provider in an SDK client. (Claude Platform OpenTelemetry docs)
• Traces are not guaranteed to be complete. Anthropic describes spans as best-effort and points to the Usage API for API-key usage tracking. (Claude Platform OpenTelemetry docs)

What Claude Platform Traces Cover

The beta turns Claude Platform activity into OpenTelemetry spans that an OTLP-compatible destination can receive. Anthropic says the launch covers Messages API requests, Claude Managed Agents, and Claude SDKs; its docs break the coverage down into API and SDK tracing. (Anthropic launch article; Claude Platform OpenTelemetry docs)

The documented traces can represent Messages API and Sessions API requests, token usage, model runs, Managed Agents turns, and tool activity. SDK tracing also covers client tools run by the tool runner and tools in a self-hosted sandbox. The exact spans depend on whether you instrument an SDK client, configure API tracing, or use both. (Claude Platform OpenTelemetry docs)

Anthropic says Claude traces follow OpenTelemetry GenAI semantic conventions. That makes it possible to inspect Claude activity alongside data from other providers in compatible tools and dashboards, rather than only inside a Claude-specific console. (Anthropic launch article)

Choose the Trace Content Mode

Metadata-only is the default mode. Anthropic's documentation says those spans can include identifiers, timing, status, token counts, and selected descriptive text such as agent names, tool names, and MCP server names. (Claude Platform OpenTelemetry docs)

Full-content mode can also include model prompts and responses. Before enabling it, check the destination's access controls and retention settings, because exported content is stored under the observability platform's policies. Anthropic also notes that organizations with certain HIPAA, FERPA, or K-12 data-handling restrictions are limited to metadata-only for Console destinations. (Claude Platform OpenTelemetry docs)

Set Up an Observability Destination

For API tracing, add a destination in the Claude Console; Anthropic says the destination must be configured by an admin. For SDK tracing, install the relevant Claude SDK and register a tracer provider. You can configure SDK tracing, API tracing, or both, depending on whether you need to observe calls from an SDK client, any API client, or a combination. (Claude Platform OpenTelemetry docs)

The launch article names observability destinations and tools including Grafana Cloud, Honeycomb, Sentry, Splunk, Datadog, Langfuse, LangSmith, New Relic, and PostHog. Anthropic's docs state the broader requirement: the destination must accept an OTLP endpoint. (Anthropic launch article; Claude Platform OpenTelemetry docs)

Treat Spans as Observability, Not Accounting

Anthropic describes exported spans as best-effort. Spans can be lost, and API callers can omit request spans by disabling sampling in the traceparent header. The documentation recommends the Usage API for tracking API-key usage, including requests that do not have spans. (Claude Platform OpenTelemetry docs)

That distinction matters operationally: OpenTelemetry can help you understand a request's path, model activity, and tool behavior, but it should not be treated as a complete ledger of API usage. For a new integration, start with metadata-only, confirm that the right spans reach the destination, then decide whether the additional detail from full-content mode is necessary. (Claude Platform OpenTelemetry docs)

Bottom Line

Claude Platform's public-beta OpenTelemetry support brings API, SDK, and Managed Agents traces into existing observability stacks. Start with the default metadata-only mode, add a destination or SDK tracer, and use the Usage API separately when you need API-key accounting. (Anthropic launch article; Claude Platform OpenTelemetry docs)

Sources

• Anthropic, Trace agents with OpenTelemetry on the Claude Platform, October 14, 2026.
• Anthropic, OpenTelemetry for the Claude API.
• ClaudeDevs, OpenTelemetry public-beta announcement, October 14, 2026.
• ClaudeDevs, OpenTelemetry trace content modes, October 14, 2026.