Anthropic introduced Claude Haiku 5.5 on 7 October as its new small model for fast, repetitive work. The input rate starts at 0.10 dollars per million tokens, but that figure alone does not describe the full cost: output has a different rate, and prompts exceeding 100,000 tokens fall into a more expensive tier.
For requests of up to 100,000 input tokens, the announced price is 0.10 dollars per million input tokens and 0.50 per million output tokens. Above the threshold, the rates increase to 0.50 and 2.50 dollars, respectively. These are API prices in dollars; they are not equivalent to the cost of a Claude app subscription.
An example that allows for comparison
A hypothetical request with 1,000 input tokens and 200 output tokens would cost 0.0002 dollars at the short-prompt tier rate, with no other tools or charges: 0.0001 for the input and the same amount for the output. The calculation illustrates the formula; it does not guarantee that a thousand-word article will use a thousand tokens. Tokens and words are not equivalent units.
Anthropic also publishes separate rates for cache reads and writes. Reusing content can change spending, but this requires the application to take advantage of that feature and its conditions to be met. A system that repeats full instructions without using a cache does not automatically get those savings.
What the company is proposing
The model is geared toward summaries, classification, queries, and other frequent, narrowly scoped tasks. It includes an adjustable effort level, which makes it possible to change the balance among speed, cost, and reasoning. This flexibility can be useful in applications that receive many small requests, provided quality is measured against real examples of the expected work.
In its tests, Anthropic reports improvements over Haiku 4.5. The published figures are the company’s results using a specific configuration; they are not tests conducted by evovo or a guarantee that the model will perform equally well on every task in Spanish. A good benchmark result also does not eliminate the need to check for data and formatting errors.
The price per token is not the whole story about efficiency
The company itself warns that it changed the tokenizer and that the same task may use slightly more tokens. For comparison, it is useful to measure the cost per useful result, including retries and corrections. Saving on the first call does little good if it has to be repeated several times afterward.
Haiku 5.5 is available on Anthropic’s platform and through the providers announced by the company. For a project in Mexico, the budget should account for the actual service bill and observed usage. The new release offers a less expensive option for certain jobs; the final choice depends on accuracy, timing, and volume, not just the headline price.
