Haiku 5.5: AI for thousands of small tasks

The part of AI you do not see

A shop receives one question about a parcel and another about a return. Before answering, someone has to decide where each enquiry belongs. That small job, repeated many times, helps explain Haiku 5.5, announced by Anthropic on 7 October. It targets classification, extraction and short tasks. Developers are the immediate audience, but the release also helps explain what might sit behind a button that sorts enquiries or summarises documents. Not every AI use needs a long conversation: sometimes the useful result is a small piece of information that lets the next step proceed.[1]

What the specification tells you

Its documentation lists text and image inputs, text output and a context capacity of up to one million tokens. A token is a unit used to divide information; it is not necessarily a whole word. The specification also allows adjustable reasoning effort. These describe capabilities, not guaranteed correct answers. Being able to supply a long document does not mean every exception inside it will be found. Readers need to ask which fact they want and how they can verify it; developers also need to consider what happens when the answer is wrong.[3]

The rate depends on each request

Claude API publishes two tiers, in dollars per million tokens. For requests with up to 100,000 input tokens: $0.10 for input and $0.50 for output. Above that threshold: $0.50 and $2.50 respectively. The threshold also counts cache reads and writes, and the higher rate applies to that entire request. In a hypothetical example, one million input tokens and 200,000 output tokens spread across short requests within the first tier total $0.20 in base costs. This is not an actual bill and excludes tools, caching and taxes. Putting everything into one long request changes the calculation.[2]

An imaginary shop, a necessary check

Return to the fictional shop. It could ask the model to label each message as shipping, returns or another enquiry, leaving uncertain cases for a person. Before using it, ambiguous messages would be useful test cases: “I want to return it because it never arrived” does not fit neatly into one category. Reviewing mistakes matters as much as celebrating quick answers. Anthropic’s model-selection guide recommends evaluation against the task. That guidance comes from the provider; we have not measured accuracy, speed or business savings here.[4]

The API and your chat subscription are different

These rates cover access for integrating the model into applications, not the monthly chat price. Anthropic announces API credits for certain Max and Team subscribers, with separate conditions and activation described in its help centre. That does not mean every user has credits or that their subscription price has fallen. If you want to chat in the app, check what your plan includes. If you are building a tool, examine consumption per request and its limits. This release adds an option for frequent tasks; its usefulness depends on the work and on tests we have not carried out here.[5]