Claude Haiku 5.5 replaces Haiku 4.5, but Priority Tier does not come with it

Claude Haiku 5.5 replaces Haiku 4.5, but Priority Tier does not come with it

Anthropic launched Claude Haiku 5.5 on October 7 as the cheaper, faster successor to Claude Haiku 4.5, and one thing does not come across with it. The migration guide tells any organisation holding a Priority Tier capacity commitment on Haiku 4.5 to "plan capacity separately: Priority Tier is not supported on Claude Haiku 5.5."

Anthropic diagram showing three layers of agent safeguards: in the model, around the model, and last resort

Anthropic's diagram of its three layers of safeguards: in the model, around the model, and last resort. Haiku 5.5 runs classifiers that can decline a request with no server-side fallback. Image: Anthropic Image: Anthropic

That is a paid guarantee on the old model with no equivalent on the new one, and it sits alongside a long list of things that have to change before a request written for Haiku 4.5 will run. Anthropic's platform release notes say it without decoration: "Code written for Claude Haiku 4.5 can break on Claude Haiku 5.5." All of these pages were read at 22:02 UTC on October 7.

Five requests that worked on Haiku 4.5 now return a 400

The migration guide numbers its checklist to ten, and Anthropic's what's new page marks five of the changes Breaking and four Changed. The five that return an error:

  • A manual thinking budget, thinking: {"type": "enabled", "budget_tokens": N}, returns a 400 error. Adaptive thinking and the effort parameter replace it.
  • Non-default temperature, top_p and top_k values return an error, and have to be omitted.
  • Assistant message prefill returns an error, so a request has to end on a user turn.
  • Computer use needs the computer_toolset_20260801 toolset in place of computer_20250124, on the Claude API and Google Cloud.
  • Editing an earlier turn invalidates thinking blocks, so conversations that send thinking back have to stay append-only.

One item on the list is a new behaviour rather than a broken one, and it also needs code. Haiku 5.5 "runs safety classifiers that can decline a request, and it has no server-side fallback", so a client has to handle stop_reason: "refusal" itself rather than relying on the API to route around it.

Thinking blocks now arrive where code expects text

Adaptive thinking is on by default, so a reply can open with one or more thinking blocks even when the request never asked for any. Code that treats the first content block as the answer gets reasoning instead of text, and Anthropic's instruction is to select blocks by their type field rather than by position. Thinking text is omitted by default unless thinking.display is set to "summarized".

A third change is a measurement problem rather than a code one. Haiku 5.5 shares the tokenizer used by Claude 4.7 and later models, and the guide says the same text counts as roughly 30% more tokens on it than on Haiku 4.5, so a max_tokens limit tuned on the old model can cut equivalent output short. The fourth binds thinking blocks to the account that made them: they "work only in the account that produced them, or in an account linked to it", which bites anyone replaying stored conversations through a second account. Anthropic applied the same binding to Claude Sonnet 5.5 on September 28.

Two groups get off lightly. Claude Managed Agents users need "no changes beyond updating the model name", and Anthropic offers everyone else a Claude Code command to do the work, /claude-api migrate this project to claude-haiku-5-5, which it says applies the ID swap and the parameter changes across a code base and then hands back a list of items to check by hand.

There is no deadline attached. Anthropic still lists Claude Haiku 4.5 at $1 and $5 per million tokens, so an organisation that wants to keep its Priority Tier capacity can stay where it is, and pay ten times the new model's short-prompt rate for the privilege. Anthropic puts Haiku 5.5's own retirement at "not sooner than October 7, 2027".

Leave a Comment

Your email address will not be published. Required fields are marked *