Anthropic Releases Claude Haiku 5.5, Cutting Costs for Repetitive Work by About 75%
Anthropic has released a small AI model aimed at high-volume, repetitive work. It is available on AWS and, for the first time in the Haiku line, lets users adjust how much effort the AI puts into each task.
Anthropic unveiled Claude Haiku 5.5, a small AI model, on October 7 local time. Amazon Web Services (AWS) also announced that it had brought the model to Amazon Bedrock and Claude Platform on AWS. Anthropic said the model costs roughly 75% less on average than its predecessor, Claude Haiku 4.5.
What changes for repetitive work
The main use case is work that has to be handled quickly and many times over, such as summarizing and classifying documents. AWS cited examples including pulling key information out of short- to medium-length documents and answering simple questions based on internal company materials. Anthropic described the model as the fastest and most efficient in the Claude 5.5 family.
A feature that lets users adjust how much effort the AI puts into each task has also arrived in the Haiku line for the first time. Instead of applying the same setting to every job, users can decide the balance between cost and problem-solving ability task by task, according to AWS. Simple classification and multi-step work can each get their own setting.
The ability to operate a computer screen directly was another focus of the upgrade. According to AWS, the model can process high-resolution images and carry out repetitive work in browsers and on the desktop. The Decoder reported that on OSWorld, a benchmark that measures computer use ability, the score rose from the previous model’s 15.7% to 72.4%.
How to divide work between Haiku 5.5 and Opus 5.5
AWS laid out a pattern for splitting duties with Claude Opus 5.5. Opus 5.5 sets the approach to complex problems and handles the hard judgment calls, while Haiku 5.5 executes clearly scoped jobs. The idea is to deploy several AI agents that work on their own and divide tasks among them, including document classification and summarization and small edits across many files.
Image: Amazon Web Services
The Decoder reported that the price of tokens, the fragments of text that models read and write, has dropped by as much as 90%. It added, however, that the way text is split into tokens has changed, so the same task consumes more tokens. That trims some of the benefit of the lower unit price.
How to use it on AWS
Amazon Bedrock users can pick the model in the console’s playground and try it right away. AWS said customers can keep using their existing access controls, usage logs and monitoring, with usage billed through AWS. On Claude Platform on AWS, users can tap Anthropic’s platform features and developer environment alongside AWS authentication and billing.
Image: Amazon Web Services
At launch, Bedrock support covers the US, Europe, Australia and Japan regions, plus a global request routing option. Claude Platform on AWS is listed as available in North America. For teams in South Korea, the questions to weigh are the quality of Korean-language work, actual billed cost and regional availability. The release lowers the cost of frequently repeated work and widens the options for separating hard judgment from simple execution.