What is Claude Haiku 5.5, and what is a small, cheap model good for?

Yesterday, Wednesday, October 7, 2026, Anthropic released Claude Haiku 5.5, which it describes as its fastest and least expensive small model yet, and its most capable one. It completes the Claude 5.5 family, after Opus 5.5 and Sonnet 5.5 last month; The New Stack notes that it is Anthropic's third 5.5 model in about a month.

Haiku has always been the smallest of Anthropic's three model sizes. Small models trade some depth of reasoning for speed and low cost, which makes them the right tool for jobs done thousands or millions of times a day, where a bigger model would be slow and expensive.

What Anthropic released

Haiku 5.5 is available to developers on the Claude Platform as claude-haiku-5-5, and on the AWS, Google Cloud and Azure clouds. Anthropic aims it at:

quick, repetitive work such as summaries, database queries and classification, sorting items into categories like "complaint," "question" or "spam";
compaction, shrinking a long conversation into a shorter summary so an AI agent can keep working without running out of memory;
subagent work, in which a larger model such as Opus 5.5 or Sonnet 5.5 leads a project and hands small, well-defined pieces to Haiku;
speed-sensitive tasks such as live customer support and browser use, since Anthropic says it is its fastest model at standard speed.

It is also the first Haiku with an effort setting, a control over how long the model thinks before answering. The New Stack reports the default is medium. Lower effort saves time and tokens; higher effort buys better answers on harder questions.

What it costs

The price change is the headline. A token is a small piece of text, about part of a word, and API prices are counted per million tokens. Haiku 5.5 has two price levels, one for short requests of up to 100,000 tokens and one for long requests above that:

Price per million	Haiku 5.5 (short request)	Haiku 5.5 (long request)	Haiku 4.5
Input	$0.10	$0.50	$1.00
Output	$0.50	$2.50	$5.00
Reading cached text	$0.01	$0.05	$0.10

That is a 90% cut for shorter requests and 50% for longer ones. According to Anthropic, nine in ten Haiku 4.5 requests fell in the shorter group. It also notes that Haiku 5.5 uses a new tokenizer, the component that splits text into tokens, which produces slightly more tokens for the same text. Taking all that into account, the company estimates Haiku 5.5 costs about 75% less to run on average.

A worked example makes the scale clearer. Imagine a shop that classifies 100,000 customer emails a month, each about 1,000 tokens in and 50 tokens out. That is 100 million input tokens and 5 million output tokens:

	Input cost	Output cost	Total per month
Haiku 4.5	$100	$25	$125
Haiku 5.5	$10	$2.50	$12.50

These are list prices for this one illustration; real bills depend on how many tokens each task actually uses.

How well it performs

On Anthropic's own tests, the jump from Haiku 4.5 is large. On the offline set of OSWorld 2.1, a test of operating a real computer through long tasks, Haiku 5.5 scores 72.4% against 15.7% for Haiku 4.5 and 48.9% for OpenAI's small GPT-6 Luna. On Terminal-Bench 4.0, a command-line coding test, it reaches 39.2%, where Haiku 4.5 scored zero.

Anthropic is clear about the limits: for complex coding, its larger Sonnet and Opus models are still the better choice. The New Stack adds that Anthropic compares Haiku 5.5 only with its own models and GPT-6 Luna, while many developers also weigh small models from companies such as Z.ai and Alibaba.

On safety, Anthropic reports far fewer misaligned behaviors than Haiku 4.5. Its cybersecurity safeguards are tighter than Haiku 4.5's and still block penetration testing, though they permit more kinds of defensive security work than the Sonnet 5.5 rules do.

The other price changes

Anthropic made two more announcements the same day. It cut the price of cache reads on Sonnet 5.5 in half, to $0.10 per million tokens. Cache reads are the cheap re-use of text a program already sent, such as long standing instructions, and Anthropic says the cut makes Sonnet 5.5 about 20% cheaper on most agent tasks. It is also adding a monthly API credit for Claude Max and Team subscribers this week: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team, usable on any model.

A worked prompt

Small models do best with narrow, explicit instructions and a fixed answer format:

Prompt

Classify the customer email below into exactly one category:
BILLING, SHIPPING, PRODUCT_QUESTION, COMPLAINT, OTHER.
Reply with the category name only, then one short reason.

Email: "My parcel says delivered but nothing arrived yesterday."

A fixed list and a one-line answer keep the output short, cheap and easy for a program to read.

Try it: Write ten short sample messages of your own and ask Haiku 5.5 to sort them with a prompt like the one above. Count how many it gets right, then try again at a higher effort level and compare.

When should you use Claude Haiku 5.5 instead of a bigger model?

Choose Haiku 5.5 when the task is narrow, repeated many times, and easy to check: tagging, routing, extracting a few fields, short summaries, or answering questions about one document. It also suits the "helper" role, where a bigger model plans the work and sends Haiku the many small lookups. And it fits anything where speed matters to a waiting person, like a support chat.

Move up to Sonnet 5.5 or Opus 5.5 when the work needs long reasoning, careful judgment or complex code. A practical approach is to start with Haiku, measure its accuracy on a sample of real cases, and send only the hard ones to a larger model.