Claude Haiku 5.5 is out: what changed, and how to prompt it well

Anthropic released Claude Haiku 5.5 on 7 October 2026. It replaces Haiku 4.5 as the small model in the Claude line-up, sitting below Sonnet 5.5 and Opus 5.5. If you use Claude through the API, Bedrock, Google Cloud or Microsoft Foundry, it's worth knowing what it's for and how prompting it well is different from prompting the bigger models.

What Haiku 5.5 is actually for

Anthropic's own documentation describes Haiku 5.5 as built for "high-volume, latency-sensitive tasks such as classification, extraction, and routing". That's a different job from Opus or Sonnet. You don't reach for Haiku when a task needs judgement or several steps of reasoning. You reach for it when you have a lot of small, well-defined tasks and you want each one back fast and cheap.

Typical jobs it suits:

• Sorting support tickets or reviews into categories.
• Pulling specific fields out of messy text (a date, a price, a name).
• Deciding which of several tools or sub-agents should handle a request.
• Quick rewrites, translations, formatting and summaries.

The numbers, from Anthropic's own documentation

Haiku 5.5 is the first Haiku model with the effort parameter that Opus 5.5 and Sonnet 5.5 already use. You can turn it up for a slightly harder batch job, or leave it at the default for routine tasks. It also uses the same tokenizer as Claude 4.7 and later models, so the same amount of text now counts as roughly 30% more tokens than it did on Haiku 4.5. That matters if you're sizing prompts close to a limit, less so for everyday use.

How this changes your prompts

1. Give it one job, not a loose brief

Haiku is not the model for "help me think through this". It does best with a single, well-defined instruction and a clear output shape. Compare:

Weak:

Look at this list of customer messages and tell me what's going on.

Better:

Classify each message below into exactly one category: Billing, Bug report, Feature request, or Other.

Return a table with two columns: the message number and the category. Do not explain your choice.

Messages:
[PASTE MESSAGES, ONE PER LINE]

The second version removes any room for Haiku to wander into analysis you didn't ask for, which is where small models waste time and tokens.

2. Say exactly what the output should look like

Because Haiku is often doing the same task hundreds of times (sorting, extracting, formatting), an unclear output format causes more damage at scale than it would with one request. Specify the exact structure: a table, a fixed set of labels, a line count, or JSON keys if you're calling it from code.

3. Use the effort parameter instead of asking it to "think more"

If a batch of tasks is slightly harder than usual, such as extraction from messy, inconsistent text, raise the effort level rather than padding the prompt with "think carefully" or "be thorough". Anthropic's guidance for the newer Claude models is the same across the family: effort controls depth more reliably than wording does.

4. Route, don't reason

If you're building something with sub-agents or multiple tools, Haiku is a sensible model to put in front: let it decide which specialised tool or bigger model should handle a request, rather than having it attempt the hard task itself. Keep its own job narrow: "pick one of these five categories" or "is this a billing question or not", not "solve this".

5. Know when to switch models

Haiku's limits show up fast once a task needs multiple steps of judgement, nuance, or weighing trade-offs. If you notice it giving a confident but shallow answer to something that actually needed more thought, that's the sign to move the task to Sonnet 5.5 or Opus 5.5, not to rewrite the prompt again. See which Claude model to use for the fuller comparison.

Prompts that suit Haiku 5.5

These are the kind of narrow, repeatable tasks Haiku is built for. The citation formatter turns a messy source list into a clean, consistently styled reference list, one of the most repetitive jobs in writing. The plain-English to SQL prompt is another good fit: a single, well-scoped question with a clear answer shape.

Other strong matches from the library: writing alt text for images, which is short and repetitive by nature, translating text while keeping the tone, and planning how to clean a messy dataset once you've broken the job into clear steps.

A copy-ready template for Haiku-sized tasks

Task: [ONE CLEAR INSTRUCTION, e.g. "classify", "extract", "rewrite in X words"].
Input: [PASTE THE TEXT OR DATA].
Output format: [EXACT SHAPE, e.g. "a table with columns A and B" or "one word: Yes or No"].
Rules: [ANYTHING TO ALWAYS OR NEVER DO]. If something doesn't fit, write "unclear" instead of guessing.

Fill it in once, and it should hold up whether you run it ten times or ten thousand.

Where to go from here

If you're deciding between Claude models rather than just prompting Haiku, see which Claude model you should use for the full comparison with Sonnet, Opus and Fable. The Claude prompts collection has hundreds of ready-made prompts, and the prompt builder can put one together for your exact task.

Quick checklist

• Give Haiku one clear, narrow task, not an open-ended brief.
• Spell out the exact output format: a table, a label, a word count.
• Raise the effort level for harder batches instead of adding "think carefully".
• Use it to route or classify, not to reason through a decision.
• Move to Sonnet or Opus the moment a task needs real judgement.

Prompts to try

Text & chat2 blanks to fillFormat references in APA, MLA or HarvardChatGPT · Gemini · Claude Text & chat2 blanks to fillWrite an SQL query from plain EnglishClaude · ChatGPT · Gemini Text & chat2 blanks to fillWrite helpful image alt textChatGPT · Gemini · Claude Text & chat4 blanks to fillTranslate while keeping tone and meaningChatGPT · Gemini · Claude Text & chat2 blanks to fillPlan how to clean a messy datasetChatGPT · Gemini · Claude

Browse all 1611 prompts →