Claude Haiku 5.5 launches with lower prices and adjustable effort

What changed

Anthropic introduced Claude Haiku 5.5 on October 7. The company positions it for frequent, narrowly scoped work such as summaries, classification, database queries, and supporting tasks inside larger agent workflows. It reports lower running costs than Haiku 4.5 and publishes performance comparisons with the release.

Those performance and savings claims come from Anthropic. Its announcement also distinguishes the model’s role from more complex agentic coding work, for which it points users toward larger models.

Source 1

The price changes at 100,000 prompt tokens

Haiku 5.5 has separate price tiers based on prompt length. The higher input and output rates apply when a prompt exceeds 100,000 tokens. Anthropic’s pricing documentation also lists separate caching charges and batch discounts; partner-operated cloud services have their own pricing.

Our analysis: a workflow that repeatedly appends history can move into a different cost tier even when each new user message is short. Budgeting from the advertised starting rate alone can therefore miss a material part of the bill.

Source 3

Where a smaller model fits

A workplace automation often contains several different jobs. Sorting incoming requests, pulling a field from a document, and deciding how to handle an exception need not share the same requirements.

Our interpretation is that the release invites teams to examine those steps separately. If a narrowly defined task is easy to verify, it may be a useful place to compare a smaller model. That is a hypothesis to test, rather than a reason to replace an existing workflow wholesale.

Measure accepted results

Start with a sample that includes ordinary cases, ambiguous cases, and missing information. Define what a correct answer looks like before comparing models. For a document extraction task, a correct value without a supporting location may still fail the requirement.

Track the cost of the initial call, retries, and any escalation to another model. Count how many outputs pass your check. Dividing the total cost by that count gives a more useful comparison than looking at the price of a single call.

For nontechnical teams, this can be a simple spreadsheet. Developers can make the same examples repeatable tests. In either case, keep the cases that fail: they reveal which work needs a different approach.

The adoption question

Haiku 5.5 is worth evaluating where the task is frequent, bounded, and checkable. Whether it saves money in a particular workflow depends on the quality of its results and what happens after an error. We have not measured those outcomes ourselves, so this article offers an evaluation path rather than a performance verdict.