Plain-English Model Explainer
What is Claude Haiku 5.5? Anthropic's fastest model, explained
The small model you reach for when the job is big in volume and small in drama.
You saw the name Claude Haiku 5.5 and want to know what it is, who makes it, and whether it matters for your work. This page covers what Anthropic says the model is for, how it differs from Haiku 4.5, how it sits next to Sonnet 5.5, Opus 5.5 and Fable 5.1, and which prompts suit it best.
Claude Haiku 5.5 is a large language model made by Anthropic and released on October 7, 2026. Anthropic calls it its fastest model to date, and the fastest and most capable small model it has released. It is designed for high-volume, latency-sensitive work such as classification, routing, extraction, summaries, and subagent tasks. It is also the first Haiku-class model with an adjustable effort setting. Its API model ID is claude-haiku-5-5.
Published
Free lesson · no account neededMeet ClaudeWhy Claude?6 minStart the free lessonKey Takeaways
- Claude Haiku 5.5 is made by Anthropic, was released on October 7, 2026, and is described by Anthropic as its fastest model to date.
- It targets high-volume, latency-sensitive work: classification, routing, extraction, summaries, compactions, database queries, and live customer support.
- It is the first Haiku-class model with an adjustable effort setting, so you choose whether to lean toward cost or toward intelligence.
- Anthropic suggests it as a fast subagent alongside Opus 5.5 or Sonnet 5.5 on coding work, with the bigger model planning and Haiku doing the quick parts.
- Haiku 4.5 is no longer in Anthropic's current lineup. It remains available on the Claude API as an older model, and Haiku 5.5 is the current Haiku.
Claude Haiku 5.5 at a glance
| Attribute | Claude Haiku 5.5 |
|---|---|
| Made by | Anthropic |
| Release date | October 7, 2026 |
| Anthropic's positioning | Fastest model to date; fastest and most capable small model Anthropic has released |
| Designed for | High-volume, latency-sensitive work such as classification, routing, extraction, and subagent tasks |
| Context window | 1M tokens |
| Maximum output | 128K tokens |
| Input and output | Text and image input, text output |
| Thinking | Adaptive thinking on by default, steered with the effort setting |
| Default effort | Medium |
| Reliable knowledge cutoff | June 2026 |
| API model ID | claude-haiku-5-5 |
| Where it is available | Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS |
| Earliest retirement | Not sooner than October 7, 2027 |
Copy and Paste
Prompts You Can Use Today
Summarize a long thread into a decision log
You are helping a project manager catch up on a long email thread. Summarize the thread below into three parts: decisions made, open questions, and action items with the person responsible. Keep each item to one sentence and do not add anything that is not in the thread. [paste thread]
Swap in a meeting transcript, a support ticket history, or a long chat export. Keep the three-part format so repeated runs stay consistent.
Sort incoming messages into categories
Classify each customer message below into exactly one category: billing, bug report, feature request, account access, or other. Return a table with the message number, the category, and a one-line reason. If a message fits two categories, choose the one the customer seems most urgent about. [paste numbered messages]
Replace the category list with your own labels. Fewer, clearly different labels give more consistent sorting.
Extract structured fields from messy text
Read the invoice text below and extract these fields: vendor name, invoice number, invoice date, due date, currency, and total amount. Return only a JSON object with those keys. If a field is missing from the text, set its value to null instead of guessing. [paste invoice text]
Change the field list to match your documents. The instruction to use null for missing values keeps the model from filling gaps.
Draft a first reply to a support question
You are a support agent for a small online software company. Write a short, friendly reply to the customer message below in under 120 words. Acknowledge the problem, give the next step in plain language, and ask for one specific piece of information if you need it. Do not promise refunds or dates. Customer message: [paste message]
Add your real tone guidelines and a few approved phrases. A person should still review replies before they are sent.
Route a request to the right team
You triage internal requests for a company with four teams: IT, HR, Finance, and Facilities. Read the request below and answer with one line in this format: TEAM | urgency (low, normal, or high) | a ten-word summary. Choose only from the four teams listed. Request: [paste request]
Rename the teams and urgency levels to match your organization. A strict one-line format makes the output easy to parse.
Compact a long working session into notes
Below is a long conversation from a working session. Compact it into a briefing that I can paste into a new chat to continue the work. Include the goal, what has been decided, what has been tried and rejected, the current state, and the next three steps. Keep it under 300 words. [paste conversation]
Compaction is one of the jobs Anthropic names for Haiku 5.5. Adjust the word limit to how much context you want to carry over.
Turn rough notes into a clean table
Turn the rough notes below into a Markdown table with the columns: item, owner, due date, and status. If the owner or due date is not stated, write 'unassigned' or 'no date'. After the table, list any lines you could not place in a row. [paste notes]
Change the columns to suit your tracker. The closing list of unplaced lines shows you what the model could not fit.
What it is built for
The jobs Anthropic designed Claude Haiku 5.5 to do
Anthropic's own wording keeps coming back to the same idea: work that is frequent, repetitive, and sensitive to waiting time.
Summaries and compactions
Anthropic names summaries and compactions as quick, repetitive workloads Haiku 5.5 handles reliably. A typical example is condensing a long thread or a long working session into a short briefing.
Classification and routing
Sorting messages into categories and sending each to the right place is central to the model's description in Anthropic's docs. These tasks arrive in large numbers, so speed adds up.
Extraction and database queries
Pulling named fields out of messy text is listed in the documentation, and database queries appear in the announcement. Both reward a model that answers quickly and consistently.
Live customer support
Anthropic says the model's speed suits speed-sensitive work like live customer support. A person waiting for a reply notices every extra second, which is where a fast model matters.
Browser use
Browser use is the other speed-sensitive example Anthropic gives. An agent that clicks, reads, and decides many times in a row feels the model's response time on every step.
Subagent work on coding tasks
Anthropic says Haiku 5.5 pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work. A bigger model leads, and Haiku handles the quick side tasks.
The plain-English version
What Claude Haiku 5.5 actually is
Claude Haiku 5.5 is a large language model, which means it is software trained to read text and images and write text in reply. Anthropic, the AI company behind the Claude family, released it on October 7, 2026. You can talk to it the way you would talk to any Claude model, and developers can call it by the API model ID claude-haiku-5-5.
Anthropic sorts its models by size and role, and Haiku is the small, quick end of that range. The company describes Haiku 5.5 as its fastest model to date and the fastest and most capable small model it has released. Small here does not mean toy. It means the model is shaped for speed and volume rather than for the hardest open-ended reasoning.
A few specifications from Anthropic's documentation show how much room it has to work with. The context window is 1M tokens, which is the amount of text the model can consider at once. A token is a small chunk of text, roughly a short word or part of a longer one. The maximum output is 128K tokens, so it can write long answers when you need them. It accepts text and image input and replies in text.
Two more details matter for how it behaves. Adaptive thinking is on by default, which means the model decides how much to reason before answering rather than always doing the same amount. And its reliable knowledge cutoff is June 2026, so anything after that date is something you should supply in your prompt instead of expecting the model to know it.
Effort setting
The new dial: what the effort setting does
Haiku 5.5 is Anthropic's first Haiku-class model with an adjustable effort setting. In Anthropic's words, this lets you decide whether to optimize for cost or for intelligence. A lower setting leans toward being quick and economical. A higher setting gives the model more room to think before it answers.
The default effort for Haiku 5.5 is medium, and adaptive thinking is on by default. Together, these mean you can use the model out of the box without touching anything, and then adjust only if a task calls for it. The effort setting is how you steer the depth of thinking that adaptive thinking applies.
In practice, this is a tuning decision you make per task type, not per message. A pipeline that tags thousands of support tickets by topic is a good candidate for a lighter touch, because each decision is simple and the volume is large. A task where the model must follow several conflicting constraints at once may be worth a higher setting. The sensible way to choose is to run a small sample of your real inputs at two settings and compare the results yourself.
Because the setting is new to the Haiku line, older advice about small models may not describe what you can do now. If you used an earlier Haiku and found it too shallow for a certain job, it is reasonable to test that job again with a higher effort setting before deciding it needs a larger model.
Compared with Haiku 4.5
What changed from Claude Haiku 4.5
Claude Haiku 4.5 was released on October 15, 2025. It is no longer in Anthropic's current lineup. Anthropic's documentation lists it with the status Active (legacy), which means it remains available on the Claude API as an older model. Haiku 5.5 is now the current Haiku model.
The first change is simply position. If you are choosing a Haiku today, Anthropic's current lineup points you to Haiku 5.5. The second change is the one Anthropic highlights directly: Haiku 5.5 is the first Haiku-class model with an adjustable effort setting. That gives small-model work a dial that the earlier Haiku did not offer.
Anthropic also describes Haiku 5.5 as the fastest and most capable small model it has released. That is a statement about the whole Haiku line to date, and Haiku 4.5 is part of that history. We are not quoting benchmark scores or percentage gains here, because the page you should trust for exact measurements is Anthropic's own documentation, and we would rather point you there than paraphrase numbers.
If you have an existing project built on Haiku 4.5, nothing in Anthropic's lineup says it stops working today. But the model documentation does say older models can be retired over time, and Anthropic maintains a model deprecations page for exactly that reason. A sensible habit is to test your real prompts on Haiku 5.5 now, so a move is a planned task rather than a rushed one.
Choosing a model
Which Claude model should you use?
Anthropic's current lineup has four models, and its documentation gives each a one-line job. Here is that positioning, in plain words.
Claude Haiku 5.5
Anthropic positions it for high-volume, latency-sensitive tasks such as classification, extraction, and routing. In Anthropic's comparison it is the fastest of the four.
Claude Sonnet 5.5
Anthropic calls it the best combination of speed and intelligence. Its latency is listed as fast, one step behind Haiku 5.5.
Claude Opus 5.5
Anthropic positions it for long-running agentic coding and knowledge work. Its latency is moderate, and Anthropic suggests starting here for most workloads if you are unsure.
Claude Fable 5.1
Anthropic positions it for demanding reasoning and long-horizon agentic work. It is the slowest of the four in Anthropic's comparison.
The helper pattern
Haiku as a fast helper next to a bigger model
Anthropic specifically says Haiku 5.5 pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work. A subagent is a smaller worker that a main agent hands a piece of the job to. The main model plans, decides, and keeps the whole picture. The subagent takes a narrow task, finishes it quickly, and reports back.
Take a coding session where the main model is working through a large change. Along the way there are small chores: scanning a set of files for a pattern, summarizing what a module does, or checking whether a function is called anywhere. Those chores do not need the deepest reasoning, but they do need to be quick, because the main model is waiting on the answer. Handing them to a fast model keeps the main thread moving.
The same shape works outside of code. A person drafting a long report with a larger model might hand Haiku the repetitive parts: tagging a hundred source notes by topic, extracting dates from a pile of documents, or compacting a long conversation so the larger model can pick it up fresh. The larger model keeps the judgment, and the small one carries the load.
If you work in Claude Code, Anthropic's command-line tool for coding with Claude, this is the pattern to look for. You can read more about that tool on our Claude Code page. The key idea is the division of labor: expensive thinking where thinking matters, fast answers where it does not.
Getting good results
How to write prompts for a fast, high-volume model
The strongest habit for repetitive work is to make the task narrow and the output format strict. Tell the model exactly which categories it may choose from, exactly which fields to return, and exactly how long the answer should be. A prompt that says return only a JSON object with these six keys leaves far less room for drift than one that says pull out the important details.
Give the model an exit for uncertainty. For extraction, say to use null when a value is not in the text. For classification, add an other category or ask the model to flag low-confidence cases. Without that, any model is tempted to fill every blank, and a confident wrong answer in a bulk job is harder to catch than an honest gap.
Keep your instructions stable and change only the input. When the same prompt runs thousands of times, a short, tested template beats a clever one-off. Write it once, test it on a dozen real examples, fix the places it stumbles, and then reuse it. This is a core skill in prompt engineering, and we walk through it step by step in our prompt engineering lessons.
Finally, show an example or two when the format matters. One sample input with its ideal output often does more than a paragraph of description. Place the examples before the real input, label them clearly, and make sure they are the kind of case you actually expect, including at least one awkward one.
Honest limits
Where a small, fast model is the wrong tool
Anthropic's own positioning is the best guide here. The company points Haiku 5.5 at high-volume, latency-sensitive tasks, and it points Fable 5.1 at demanding reasoning and long-horizon agentic work. If your task is a long chain of dependent decisions where an early slip ruins everything after it, that is closer to the work Anthropic describes for the larger models.
Speed also does not remove the need for checking. For anything that affects a customer, a payment, a legal position, or a medical or safety decision, a person should review the output. A draft support reply that goes out unread is a risk no matter which model wrote it. Use the model to produce the first version quickly, and keep a human in the loop for what matters.
The knowledge cutoff is another limit to plan around. Haiku 5.5's reliable knowledge cutoff is June 2026, so it will not reliably know about events, product changes, or documents that came later. For anything recent, paste the relevant text into the prompt and tell the model to answer only from what you gave it.
Finally, a model like this only helps if you measure it on your own data. A prompt that sorts your tickets well says little about someone else's tickets. Keep a small set of examples with known right answers, rerun them whenever you change the prompt or the effort setting, and treat any drop as a signal to look closer.
Access and availability
Where you can use Claude Haiku 5.5
For developers, the model ID is claude-haiku-5-5. Anthropic's documentation lists it as available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. That spread matters if your company already standardizes on one cloud provider, because you may be able to use the model without adding a new vendor relationship.
Anthropic also publishes the system prompt it uses on claude.ai and the Claude apps for this model. A system prompt is the set of background instructions a product gives the model before your conversation starts. Having it published means you can read how Anthropic frames the model's behavior in its own apps, which is useful if you are building something similar.
On lifespan, Anthropic states it will not retire Haiku 5.5 sooner than October 7, 2027. For a team building a product, that is a planning anchor: you have a stated floor on how long the model will be around, and you can schedule any future migration around it instead of guessing.
Which plans or apps include Haiku 5.5 can change, so check the model picker or Anthropic's current documentation instead of relying on a page like this one. If you are new to the Claude apps, our getting started and what is Claude pages cover the basics before you pick a model at all.
Which one for you
Which model to pick, by the kind of work you do
If your work is a steady stream of small, similar jobs, such as sorting messages, extracting fields, summarizing, or drafting short replies, Haiku 5.5 is the model Anthropic designed for that. Start there, test it on a sample, and move up only if the results are not good enough.
If you want a balance and do a mix of writing, analysis, and everyday questions, Sonnet 5.5 is Anthropic's best combination of speed and intelligence, and it is only one step slower than Haiku 5.5. If you are not sure where to begin, Anthropic suggests Opus 5.5 for most workloads, particularly long-running agentic coding and knowledge work. Our Sonnet versus Opus page goes deeper on that choice.
If your task is the hard kind, with demanding reasoning or long-horizon agentic work, Anthropic points to Fable 5.1, and you accept that it is slower. You can read more on our Fable 5.1 page.
And if you are running something large, the best answer is often more than one model. Let a bigger model plan and judge, and let Haiku 5.5 take the high-volume pieces. That is the pattern Anthropic itself describes, and it is the one worth practicing.
Learn to put the right Claude model to work
Reading about a model and being good at using it are different things. The free Nightschool AI curriculum is hands-on from the first lesson, so you practice real prompts instead of just reading about them. Start with a free lesson and build the habits behind every prompt on this page.
Frequently Asked Questions
What is Claude Haiku 5.5?
Claude Haiku 5.5 is a large language model made by Anthropic and released on October 7, 2026. Anthropic calls it its fastest model to date and the fastest and most capable small model it has released. It is designed for high-volume, latency-sensitive work such as classification, routing, extraction, summaries, and subagent tasks.
When was Claude Haiku 5.5 released?
Anthropic released Claude Haiku 5.5 on October 7, 2026. The release date is stated in Anthropic's model documentation. Anthropic has also said it will not retire the model sooner than October 7, 2027, which gives developers a stated minimum window for planning.
What is Claude Haiku 5.5 best at?
Anthropic says it is built for high-volume, latency-sensitive work: classification, routing, extraction, summaries, compactions, and database queries. Because it is fast, Anthropic also says it works especially well for live customer support and browser use, and as a subagent next to Opus 5.5 or Sonnet 5.5 on coding work.
Is Claude Haiku 5.5 the fastest Claude model?
Yes. Anthropic describes it as its fastest model to date. In Anthropic's lineup comparison, Haiku 5.5 is listed as fastest, Sonnet 5.5 as fast, Opus 5.5 as moderate, and Fable 5.1 as slower. Faster does not mean better for every task, since the larger models are aimed at harder work.
What is the difference between Claude Haiku 5.5 and Sonnet 5.5?
Anthropic positions Sonnet 5.5 as the best combination of speed and intelligence, and Haiku 5.5 as the choice for high-volume, latency-sensitive tasks such as classification, extraction, and routing. Haiku 5.5 is faster. A practical approach is to test your own examples on both and pick the faster one if the results hold up.
What changed from Claude Haiku 4.5 to Haiku 5.5?
Haiku 5.5 is now the current Haiku model, and Haiku 4.5 is no longer in Anthropic's current lineup. It stays available on the Claude API as an older model. The headline change Anthropic names is that Haiku 5.5 is the first Haiku-class model with an adjustable effort setting.
What is the effort setting in Claude Haiku 5.5?
The effort setting lets you steer how deeply the model thinks, so you can decide whether to optimize for cost or for intelligence. Adaptive thinking is on by default, and the default effort is medium. Haiku 5.5 is the first Haiku-class model to offer this setting.
What is the context window of Claude Haiku 5.5?
Anthropic's documentation lists a 1M token context window and a maximum output of 128K tokens. The model accepts text and image input and produces text output. Its reliable knowledge cutoff is June 2026, so supply any later information in your prompt.
What is the API model ID for Claude Haiku 5.5?
The API model ID is claude-haiku-5-5. Anthropic lists the model as available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. Check Anthropic's current documentation for the exact identifier format your platform uses.
Which Claude model should I use if I am not sure?
Anthropic suggests starting with Opus 5.5 for most workloads if you are unsure. Then test whether a faster model does the job well enough. For repetitive, high-volume tasks such as sorting or extraction, try Haiku 5.5. For the hardest reasoning and long-horizon agentic work, Anthropic points to Fable 5.1.
Can Claude Haiku 5.5 work alongside a bigger model?
Yes. Anthropic says Haiku 5.5 pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work. The larger model plans and decides, while Haiku handles quick, narrow tasks. The same pattern can apply to non-coding work such as tagging, extraction, and compacting long conversations.
Ready to try it yourself?
The fastest way to understand a model is to give it real work.