[FEATURE] Include thinking/extended-thinking tokens in OpenTelemetry metrics

Resolved 💬 4 comments Opened Mar 6, 2026 by JuanjoFuchs Closed Apr 7, 2026

Feature Request

Claude Code's OpenTelemetry integration exports token counts (input, output, cache read, cache creation) but does not include thinking/extended-thinking tokens.

Why this matters

Thinking tokens can be 3-10x the visible output tokens and count toward rate limits. Without them in the telemetry, any cost or token dashboard built on OpenTelemetry data undercounts what Claude actually consumes. Users building observability dashboards (like claudefana) have no way to show the full picture.

Requested change

Add thinking token counts to the token_usage metric and/or as attributes on API request log events, alongside existing input/output/cache fields.

Related

This is a re-filing of #16943 which was closed.

View original on GitHub ↗

This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗