Claude Haiku 5.5 cuts the cost of small AI tasks and makes routing more valuable

Anthropic released Claude Haiku 5.5 on 7 October for high-volume AI work. The company estimates roughly 75% lower average running costs than Haiku 4.5. The more consequential possibility is that tasks previously too small to justify a model call can become practical parts of a larger workflow.[1]

Think of a document-processing system. Finding the invoice number, identifying the currency and deciding which queue should receive the file are different jobs from resolving an ambiguous contract clause. Sending everything to the most expensive reasoning model treats these jobs as equally demanding. A capable small model creates room for a more deliberate division of labour.

The 100,000-token line is a billing boundary

The higher band applies to a request whose total prompt exceeds 100,000 tokens; it is not a surcharge confined to the excess tokens. Cache reads and cache writes also count toward that threshold. A cache hit can reduce input charges without returning a long request to the short-prompt band. Each request is priced separately. This makes retrieval and prompt length real billing decisions: attach the evidence the task needs, and count the actual request before estimating its cost.[7]

What 100,000 routine jobs could cost

Our fixed-token illustration assumes each request contains 10,000 uncached input tokens and produces 1,000 billable output tokens. At the short-prompt rates, that is $0.001 for input plus $0.0005 for output: $0.0015 per request, or $150 for 100,000 requests. This is arithmetic using published rates, not a measured Haiku workload.[1]

Now suppose every tenth item requires a specialist escalation costing an assumed additional $0.03. The 10,000 escalations add $300, making the combined model bill $450. A much cheaper first pass does not guarantee an equally dramatic end-to-end saving. Equally, a good first pass may spare the more capable model from processing ninety thousand straightforward cases. The escalation rate becomes a design variable.

The critical operational question is which items get escalated. A difficult-looking document is not necessarily a difficult task. Conversely, a short message may contain an exception that matters enormously. Routing should use observable conditions—missing fields, inconsistent totals, unsupported categories or a failed validation—not simply the assistant’s confident tone.

A tokenizer change can move the boundary

Anthropic’s migration guide says identical text produces approximately 30% more tokens with the new tokenizer than with Haiku 4.5, depending on content. It instructs developers to recount inputs and revisit output limits. A stored estimate from the previous model is therefore not a reliable new bill.[3]

For an illustrative request previously counted at 80,000 tokens, a 30% increase would mean 104,000. That would cross the published price-band boundary. It is not a prediction that every such request will cross: the exact text must be counted with the intended model. This is why comparing token prices and comparing the cost of processing the same document are different calculations.[3][1]

Caching introduces another distinct opportunity. Anthropic’s documentation describes reusing a matching input prefix, with separate usage records for cache reads and cache creation. This is not a store of previously generated answers. A stable instruction block or repeated source prefix may be reusable even when the question changes.[5]

The implication for system design is to separate stable material from changing material. Do not scramble an unchanged instruction prefix on each call and then assume it will still be reusable. At the same time, do not preserve outdated instructions merely to protect a cache hit. Correctness must determine what is reusable; accounting follows.

Migration is a behavioural change, not just a price change

The new model supports adaptive thinking and an effort setting. The migration documentation warns that older explicit thinking-budget requests can fail, and that code assuming the first returned content block is the answer needs to select blocks by type instead. Thinking also consumes the output budget. These are integration issues that a successful model-name substitution would not test.[3][4]

A narrow deployment is easier to evaluate than a wholesale replacement. Choose a meaningful task with checkable outputs: extracting a reference number, assigning one permitted label, or producing a short summary that must preserve named facts. Anthropic’s evaluation guidance supports task-specific success criteria and representative test cases rather than judging one impressive response.[6]

Keep a fixed sample of routine cases and a smaller collection of expensive mistakes. Measure both. For example, a classification system may handle ordinary messages well while mishandling a rare cancellation condition. An average score can conceal that asymmetry. Our recommendation is to make the routing policy sensitive to consequences, not only to frequency.

Haiku 5.5 is interesting because a fast, inexpensive component can change the shape of an application. More work can be screened, structured and checked before the hardest cases reach a larger model. The winning design is not necessarily the one with the smallest model bill. It is the one that buys the appropriate amount of intelligence at each step and makes exceptions visible.

Report an error or suggest a correction