Anthropic has introduced Claude Haiku 5.5, the smallest model in its Claude 5.5 family, targeting frequent tasks without the cost of larger models. The company estimates that it costs around 75% less to run on average than Haiku 4.5. The change is not only about price: for the first time in the Haiku line, developers can adjust the model's effort for each request.
For prompts up to 100,000 tokens, pricing is $0.10 per million input tokens and $0.50 per million output tokens. Above that threshold, the rates rise to $0.50 and $2.50 respectively. The lower rates match those listed for GPT-6 Luna; for longer prompts, Haiku 5.5's prices increase fivefold. Anthropic's 75% average saving over Haiku 4.5 is an estimate, not a uniform reduction on every request.
Five levels let developers balance cost and capability
Image: AWS The effort controls offer five settings: low, medium, high, xhigh, and max. The model has no option to turn reasoning off entirely and defaults to medium. Developers can therefore spend more resources on difficult requests and use a lower level for repetitive classification or summarization.
Anthropic recommends Haiku 5.5 for high-volume, cost-sensitive work: summarizing and classifying documents, extracting information, answering questions over knowledge bases, and making focused code changes. It can also act as a subagent, a model assigned simpler steps by a primary agent. In a coding workflow, for example, Haiku can route requests or perform an initial review, leaving a more capable model to handle problems that need deeper reasoning. Its speed is also intended for interactive features such as live customer support.
Results in Anthropic's published tests
Image: Anthropic In the table published by Anthropic, Haiku 5.5 scores 72.4% on the offline subset of OSWorld 2.1, a benchmark that tests computer use. Haiku 4.5 scores 15.7% on the same test. On the Humanity's Last Exam reasoning benchmark, the new model records 45.9% without tools and 57.4% with tools; Haiku 4.5 scored 10.2% and 18.7%, respectively. These are results reported by the company, not independent verification.
There is another factor to consider when estimating API costs: Simon Willison notes that Haiku 5.5's new tokenizer can count more tokens than the previous one for the same prompt. Token counts determine API costs, so the per-million rate alone does not predict the cost of a workload. Willison also points out that above 100,000 tokens, GPT-6 Luna has lower rates than Haiku 5.5.
Haiku 5.5 is available on the Claude Platform and Amazon Bedrock, including through the geographic regions listed by AWS. The model completes the Claude 5.5 family, following Sonnet 5.5, aimed at lower-cost everyday coding, and Opus 5.5 for more complex tasks.
No comments yet. Be the first!