Correction: Claude's 'thinking blocks' are summarized user-facing messages, not reasoning
Sauers_ · x · 2026-10-18
Sauers issues a correction after @smitop's pushback: what looked like "thinking" in Claude's API thinking blocks is actually connector text summarization — short user-facing messages between tool calls get rewritten by a separate summarizer model, and Claude itself never sees the summarized output.
The accompanying docs clarify the key distinction: the thinking block type carries two different things — Reasoning (internal work, shown as a summary) and Progress updates/narration (messages intended for the user, also summarized). So those 240 narration blocks were summarized user-facing commentary, not leaked private reasoning. Since Claude receives its original wording back via the encrypted signature while users see the summary, this can explain wording mismatches when asking Claude "do you remember saying this exact phrase?"
Related event: Claude's "Thinking Blocks" Revealed to Be Summarized User Messages(2 posts)→
More from Models
- GPT-6 Astra and DeepSeek V4.1 share the same tic: reasoning traces full of "Need" — scaling01 · 2026-10-18
- Frontier model defensive access locked behind un-applyable 'Daybreak Red' program — andersonbcdefg · 2026-10-18
- Frontier models obsessed with 'sealing predictions' and self-coherence, users report — repligate · 2026-10-18
- Why DeepSeek uses dated sub-versions: new major versions need fresh Chinese approval — TheZachMueller · 2026-10-18
- OpenRSI Index: a benchmark for research agents doing recursive self-improvement — jiqizhixin · 2026-10-18
- Perplexity's decision model wins Catan tournament at under half a cent per game — denisyarats · 2026-10-18