Haiku 5.5 migrations require thinking, sampling and history changes
Messages API code moving from Haiku 4.5 to Haiku 5.5 needs request and response-handling changes. Manual thinking budgets and assistant prefill fail, adaptive thinking is the default, and token counts and conversation-history rules change.
Kingy verdict
Haiku 5.5 needs a client migration for integrations that use manual thinking budgets, sampling controls or assistant prefill. Review request fields, response parsing and conversation storage together before changing the model ID. Forced tool choice remains supported, and no mandatory migration deadline is established.
Required action
Before switching a Messages API integration, audit thinking and sampling fields, block parsing, assistant prefill, tool routing and stored history. Recount prompts with the target model and revisit output limits and the applicable pricing tier. Validate the exact platform and client before production migration.
Scope and notice
Anthropic documented the change on October 7, 2026. This checklist covers Messages API code moving from Haiku 4.5 to claude-haiku-5-5. The migration guide says Claude Managed Agents users need only update the model name for this model migration. Platform model IDs differ; use the identifier documented for your provider.
Thinking and response parsing
Replace thinking: {"type": "enabled", "budget_tokens": N} with adaptive thinking; manual budgets return HTTP 400. Adaptive thinking is on by default. Select response blocks by type rather than reading the first block as the answer, and pass thinking blocks back unmodified with tool results. Thinking counts toward max_tokens, so a small limit can end before any answer text. Thinking text is empty by default; use display: "summarized" if you need the documented summarized display. Thinking can be disabled at high effort or below; disabled thinking at xhigh or max returns HTTP 400.
Sampling, prefill and forced tools
Omit temperature, top_p and top_k. If temperature is sent, only 1 is accepted; if top_p is sent, only 0.99 is accepted. Sending both, any top_k, or other values returns HTTP 400. End messages with a user turn instead of an assistant prefill. Haiku 5.5 accepts forced tool_choice, including any or a named tool: the response starts with the tool call and has no thinking block. Use auto if the model should be able to think before calling a tool. Sonnet-specific forced-tool restrictions do not apply here.
Computer use and stored conversations
On the Claude API and Google Cloud, replace computer_20250124 with computer_toolset_20260801 and follow the guide’s agent-loop and header changes. Other platforms need their own compatibility checks. Replay thinking blocks through the producing account or a linked account; another account’s request can succeed after the block is dropped. Keep system, tools and earlier messages unchanged when replaying thinking blocks. For accounts created before August 31, 2026 at 00:00 UTC, the prefix-change error applies only to requests that set thinking.block_binding.prefix_mismatch_behavior.
Token budgets and capacity
Recount prompts with the target model: the same text uses more tokens, and old max_tokens limits and cost estimates need review. The first-party pricing table separates prompts up to 100,000 tokens from prompts over 100,000 tokens; no single flat rate is assumed here. Handle stop_reason: "refusal"; server-side fallback is unavailable. Haiku 4.5 Priority Tier commitments need separate capacity planning because Haiku 5.5 does not support Priority Tier.
No scheduled retirement
The model table’s “not sooner than October 7, 2027” commitment is a lower bound, not a scheduled shutdown. This migration record creates no calendar milestone or deadline.
Evidence limits
Historical first detection is unverified. The retained October 7 source observation establishes this evidence capture, not the earliest time Kingy detected the change. These are documented behaviors. Kingy has not run an inference, account, migration, entitlement, performance or billing test for this packet.