On October 7, 2026, Anthropic released Claude Haiku 5.5, the small model for its AI assistant Claude. Haiku is the fastest and lowest-priced line among Claude's models. Anthropic positions it as a model for work that involves high volumes of processing or needs fast responses, and says it costs about 75% less on average than Haiku 4.5.
The announcement
Haiku 5.5 is the small model of the 5.5 generation, following Claude Opus 5.5 and Claude Sonnet 5.5, which were released in September. Anthropic lists tasks like the following as good fits.
- Summarizing text and compacting long conversations
- Classification, information extraction and database queries
- Subagents that handle detailed tasks under a higher-tier model
- Uses where response speed matters, such as live customer support and browser operation
The amount of text it can handle (the context window) has grown from Haiku 4.5's 200,000 tokens to 1 million tokens, and the amount it can output in one go has grown from 64,000 tokens to 128,000 tokens. It accepts text and images as input.
Pricing
API pricing is split into two tiers based on the length of each request. Prices per million tokens are as follows.
- Requests of up to 100,000 tokens: $0.10 for input and $0.50 for output (Haiku 4.5: $1 for input and $5 for output)
- Requests over 100,000 tokens: $0.50 for input and $2.50 for output
That is one-tenth of Haiku 4.5's price for short requests and half for long ones. However, Haiku 5.5 counts text using a new method, so the same text comes out to about 30% more tokens than with Haiku 4.5. How much your actual costs fall will depend on the length and content of your requests.
Performance
In benchmarks published by Anthropic, the model jumped from Haiku 4.5's 15.7% to 72.4% on OSWorld 2.1 (the tasks that can be run offline), which measures computer-use ability. On GDPval-AA v2.1, which evaluates knowledge work in real jobs, its score also rose from Haiku 4.5's 735 to 1,620. Both are figures published by Anthropic (GDPval-AA is an evaluation metric from Artificial Analysis).
The model also now supports effort settings that let you choose how deeply it thinks. You can choose from five levels, low, medium, high, xhigh and max, with medium as the default. This lets you have it think less on simple requests to be faster and cheaper, and think deeply on difficult ones.
Where to use it
Haiku 5.5 is available on the web version of Claude (claude.ai) and in the iOS and Android apps. For developers, it is offered through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS.
Developers using it through the API should watch for some changes when switching from Haiku 4.5. For example, developers are asked not to send generation settings such as temperature, and specifying a non-default value results in an error; in addition, the previous way of specifying how much the model thinks can no longer be used. Anthropic has published a migration guide.
CoAI's take
What stands out about Haiku 5.5 is that it greatly improves computer-use performance while sharply cutting prices. As splitting work across multiple AIs becomes more common, cheap and fast models that handle detailed tasks are becoming more important. Combinations such as leaving overall judgment to Opus 5.5 or Sonnet 5.5 and passing subtasks to Haiku 5.5 look increasingly practical. On the other hand, because the way tokens are counted has changed, the surest way to know your actual costs is to test with your own usage.
