What changed
On September 28, 2026, ClaudeDevs said it is extending preserved thinking with Sonnet 5.5 to curb distillation attacks that switch accounts. In that post, thinking stays in the original organization that generated it, and switching accounts mid-session makes Claude reread the session and generate new thinking.
On Anthropic's preserved-thinking page, preserved thinking decides whether the model may use a thinking block sent back from an earlier turn. A thinking block is that earlier-turn reasoning, included again with the history. Sonnet 5.5 blocks work only in the account that produced them, or in an account linked to it. The account rule leaves thinking blocks from earlier models alone.
ClaudeDevs describes the switch as a reread that produces new thinking. The docs describe a quieter result: the request succeeds, the API drops the other account's block before the model sees it, and the model answers without that reasoning. On the Claude API and Google Cloud, the thinking-binding-controls-2026-08-01 beta header makes the response list each of those drops in input_transformations, as thinking_dropped with reason organization_binding_mismatch. Without the header, the drop is silent. A dropped block is not billed.
Who should check
Check an integration that sends a Sonnet 5.5 thinking block from one account in a request made by a different account. Claude Code, claude.ai, Claude Managed Agents, and the Claude Agent SDK need no change when they build the request. Code that keeps the system prompt and tools fixed for a session and only appends messages is in that same unchanged group.
Account binding is separate from the prefix check. The prefix is the system prompt, the tools, and the messages before a thinking block. If that prefix changes, the block is no longer valid. On accounts created on or after August 31, 2026, 00:00 UTC, the API enforces this check on Fable 5.1, Opus 5.5, and Sonnet 5.5, and it returns a 400 unless prefix_mismatch_behavior is drop_block. That new-account date applies on the Claude API and on cloud platforms. On an older account, the API enforces the check only when the request sets prefix_mismatch_behavior, so a clean run on an older key can miss the 400 a newer key returns. The field needs the thinking-binding-controls-2026-08-01 beta header. Under drop_block, the dropped blocks are not billed, but Claude can think more to recreate the missing reasoning, and the session can use more tokens.
block_binding is the setting that carries prefix_mismatch_behavior. between_tools is a thinking mode. On Sonnet 5.5, block_binding is accepted only when the thinking type is adaptive. Sending block_binding with between_tools returns a 400.
Which model can read the block is a third check. Sonnet 5.5 keeps reasoning from Sonnet 5, Opus 4.8, Haiku 4.5, and earlier models. A move onto Sonnet 5.5 from Opus 5, Opus 5.5, Fable, or Mythos continues without that prior reasoning. No other model reads Sonnet 5.5 thinking, so a move away from Sonnet 5.5 to any other model, including a server-side fallback, continues with those blocks dropped.
Availability, price, and the reset
ClaudeDevs said Sonnet 5.5 is available now in the Claude Platform and Claude Code. The Claude Code changelog adds claude-sonnet-5-5 as the default Sonnet model on the Anthropic API, with 1M context, $2/$10 per Mtok, and $0.20/Mtok cache reads. The changelog writes $2/$10 as a pair and does not label input and output.
@claudeai called Sonnet 5.5 the second model in the Claude 5.5 family, more than 30% faster than Sonnet 5 and up to 30% less for most work. ClaudeDevs said it is smarter and more efficient than Sonnet 5, 30% faster, and up to 30% less for most work, so Claude Code usage goes further, and pointed it at bug fixes and quick feature iteration. ClaudeDevs also said all Pro, Max, and Team users were granted a reset last Tuesday. If it has not been used yet, it can still be applied whenever you want, before Oct 22. That post gives no credit amount, and Oct 22 has no timezone.
What is still open
The preserved-thinking page allows an account linked to the one that produced the block, and it does not say which accounts qualify. It describes the organization_binding_mismatch listing for the Claude API and Google Cloud. The reread that produces new thinking is ClaudeDevs' description of a mid-session switch. The page says the other account's request succeeds, the API drops the block before the model sees it, and the model answers without that reasoning. The reset post gives no credit amount, and its date of before Oct 22 has no timezone.
Source
Also from
Sources and editorial approach
The account rule, the prefix check, and which models can read Sonnet 5.5 thinking come from Anthropic's preserved-thinking page, checked on September 28, 2026. The Claude Code changelog, checked the same day, is the source for the line that adds claude-sonnet-5-5 as the default Sonnet model on the Anthropic API, with 1M context, $2/$10 per Mtok, and $0.20/Mtok cache reads. Availability, the speed and savings comparisons, and the usage reset come from the ClaudeDevs post and the same-day @claudeai and ClaudeDevs launch posts on September 28, 2026. The reset page at clau.de/reset was not opened. ClaudeDevs names Pro, Max, and Team users, says the reset was granted last Tuesday, and says an unused reset can still be applied before Oct 22. That post gives no credit amount and no timezone. No speed run or price comparison was performed.