Claude Haiku 5.5 (claude-haiku-5-5) became the default Haiku model with Claude Code version 2.1.293, released on 7 October. According to a Classmethod write-up of the release, it has a 1M-token context window and is priced at $0.10 per million input tokens and $0.50 per million output tokens, rising for prompts over 100,000 tokens. Check Anthropic’s pricing page before budgeting on those figures, as we have only seen them reported second-hand.
Why a small model matters
Anthropic’s Haiku line is its fast, low-cost tier. When that tier improves, tasks that used to need the flagship model start to fit on the cheap one: classifying support tickets, extracting data from documents, summarising calls, drafting replies, powering search inside an app. These are the unglamorous AI features that most products actually ship.
Two things follow for anyone commissioning an AI product:
- Running costs fall. The monthly bill for an AI feature is mostly model usage. Moving a high-volume task to a smaller model can cut that dramatically, if quality holds.
- Speed improves. Small models respond faster, which is often the difference between an AI feature that feels native and one that feels bolted on.
The catch
Cheaper does not mean automatically right. A smaller model can be confidently wrong on harder tasks, and swapping models without testing is how quality quietly drops. We saw a related risk when OpenAI retired a model on a fixed date and left hard-coded apps scrambling. Good AI products route work between model sizes, test against real examples, and make the model a setting rather than something baked into the code.
So what
If someone quoted you for an AI feature a year ago, the economics have probably shifted. Ask for a refreshed estimate that shows which tasks go to which model tier, and how quality is measured. That is the work our team does when building AI products, and we are happy to review an existing quote. Start a project to talk it through.