Claude's thinking-block 400 explained: preserved thinking, not a bug

The rule behind the error is straightforward once you see the mechanics. Each thinking block carries a signature recording the conversation prefix that produced it — the top-level system prompt, the set of tools, and every message that came before that block. When history is resent, the API checks whether that prefix is unchanged. If it is not, that block and all thinking blocks after it are treated as invalid, which is what surfaces as the opening 400.

The awkward part for builders is that the check is conditional on whether the model actually thought. In the October 8 test, claude-fable-5-1 produced no thinking blocks at all across three easy "effort low" turns, so trimming the first exchange changed nothing. claude-opus-5-5 also did not think on the first two turns, but did on the third — and that was the block that got dropped as messages.3.content.0. The practical consequence is that the same history-editing code can pass an easy test and fail on a hard production input.

Its own audit illustrates where the line falls. Three projects calling the Claude API — a subtitle-formatting pipeline, a reading-map web app, and a daily job that reads securities reports — were all single-request calls with no reassembled history, and all showed zero truncation findings. The daily job also happened to order its context as system, then PDF, then per-analysis instructions, keeping the variable parts at the end — a layout built for caching that matches the append-first remedy.

For a one-time stopgap, the API can be told to drop offending blocks rather than fail, using the thinking-binding-controls-2026-08-01 beta and prefix_mismatch_behavior set to "drop_block"; the dropped items then appear in input_transformations. But since the default differs by surface — error on headerless enforced accounts, drop_block when only the header is sent — setting the behavior explicitly is advised.