Sonnet 5.5 changes thinking, tool choice and conversation-history handling
Anthropic’s Sonnet 5.5 migration guidance changes thinking configuration, rejects forced tool choice and applies model, prefix and account conditions to thinking history. Legacy computer-tool and advisor restrictions have their own scopes; inter-tool text also changes response shape.
Required action
Before routing a workflow to claude-sonnet-5-5, audit thinking settings, tool-choice assumptions, append-only history, streaming content-block parsing, computer tools and advisor models. Validate the affected platform and effort setting.
Migration checklist
| Affected assumption | Documented Sonnet 5.5 behavior | Action |
|---|---|---|
| thinking.type=disabled | Rejected; between_tools only at low/medium/high | Use between_tools there, or adaptive thinking for xhigh/max. |
| Extra between_tools fields or per-turn effort changes | display/budget_tokens/block_binding and changed effort are rejected | Keep effort fixed; use adaptive where those controls are needed. |
| tool_choice any/tool | Returns 400; auto and none are supported | Audit code that requires a forced call; auto is not a call guarantee. |
| Editable replayed history | Prefix validation has account-age and opt-in conditions | Keep system, tools and earlier messages unchanged; inspect configured mismatch handling. |
| Thinking from another model or account | Unreadable model/account blocks are dropped; requests can succeed | Track silent loss of reasoning; retain documented compatibility and linked-account exceptions. |
| computer_20251124 | Not accepted on Claude API or Google Cloud | Follow the platform-specific computer-use migration guidance. |
| Advisor set includes Opus 4.8, Opus 4.7 or Sonnet 5 | Those advisors are rejected | Choose a documented supported advisor; verify access. |
| Inter-tool progress arrives as text | Progress arrives in thinking blocks; default display omits it | Parse by block type and review display/summary handling. |
Thinking configuration and effort
The September 28 release and Sonnet migration guide document adaptive thinking by default. disabled and manual enabled budgets are not supported. between_tools turns off up-front thinking at low, medium or high effort. It rejects xhigh/max and extra display, budget_tokens or block_binding fields; it also requires the effort level to stay fixed during the conversation.
Forced tool choice
tool_choice values any and tool return a 400 error. auto and none are supported. Strict tool schemas can help validate tool input, but auto plus strict is not a guarantee that a tool will be called. Review workflows that require a specific external action.
Three history rules with different outcomes
Model compatibility, prefix integrity and account binding are separate checks. Unsupported model blocks are dropped rather than proving a failed request. Sonnet 5.5 blocks also work only in their originating account or a linked account; another account’s request can succeed after the blocks are dropped. Earlier models’ blocks do not acquire this new account rule.
Prefix checks depend on account age and configuration
For accounts created on or after August 31, 2026 at 00:00 UTC, prefix checking is on by default on Claude API, Bedrock and Google Cloud. Earlier accounts opt in by setting prefix_mismatch_behavior. Ordinary enforced prefix mismatches default to a 400; Message Batches drops failing blocks when that field is unset. The beta controls and mismatch reporting require thinking-binding-controls-2026-08-01. Sonnet’s block_binding controls require adaptive thinking; they cannot be added to between_tools.
Preserve history without claiming all models share reasoning
Keep the system prompt, tools and earlier messages stable and append new turns. Pass thinking blocks back unchanged, including empty blocks. Model-switch compatibility is selective: Sonnet 5.5 can read Sonnet 5, Opus 4.8, Haiku 4.5 and earlier blocks, but not Opus 5/5.5 or Fable/Mythos blocks. Only Opus 5.5 on Claude API and Google Cloud is documented here to read Sonnet 5.5 thinking when switching away.
Computer tools, advisors and streaming
The older computer_20251124 tool is rejected on Claude API and Google Cloud; do not transfer that restriction to every partner. The advisor tool rejects Opus 4.8, Opus 4.7 and Sonnet 5. Progress between tool calls arrives in thinking blocks, with text omitted by default. Review response parsing and streaming display rather than assuming a quiet stream means no progress.
No fixed migration deadline
This is a model-migration record, separate from the Sonnet 4.5 November 30 retirement candidate. September 28 is the dated release notice, not an invented shutdown deadline. Kingy has not reproduced errors, tested SDK behavior or checked any account entitlement.
Evidence limits
Historical first-detection date is unverified. This packet was prepared from retained and newly retrieved official evidence; preparation and retrieval dates are not substituted for first detection.
- No vendor API, SDK, account entitlement or runtime failure was tested.
- Do not transfer Sonnet 5.5 rules from or onto Opus, Fable or Mythos by analogy.
- The prefix default threshold, older-account opt-in, Message Batches exception and required beta header remain explicit.
- Model/account mismatch drops and prefix rejection are distinct; linked-account and model-compatibility exceptions remain.
- Computer-tool rejection is limited to the documented Claude API/Google Cloud surfaces.
- No migration deadline is invented. Sonnet 4.5’s retirement candidate remains separate.