Claude Haiku 5.5: Anthropic's cheapest and fastest model

What's new

On October 7, 2026, Anthropic released Claude Haiku 5.5. It is the smallest and lightest model in the Claude 5.5 family, and it follows Opus and Sonnet from the same family. Anthropic calls it its fastest, cheapest and most capable small model yet. It is built for work that has to be repeated many times and cost little.

Haiku 5.5 is the first Haiku model with so-called effort settings. It is a kind of dial where the developer chooses how much the model should think about each task. It still has room for 1 million tokens in memory (a token is a small piece of text) and can respond with up to 128,000 tokens at a time. The model can be used via the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.

What's clever

The biggest news is the price combined with the quality. According to Anthropic, Haiku 5.5 is better than its predecessor Haiku 4.5 at programming, tool use and computer use.

Speed is also in focus. The project tool Asana, which has tested the model, says tasks were completed more than 30 percent faster than with the model it uses today. Each step in an AI agent (a program that solves tasks in multiple steps on its own) ran up to 2.5 times faster.

The effort dial is also practical. Simple tasks can be handled cheaply and quickly, while hard tasks can get more thinking time.

Cheaper or better?

Both. For API use (that is, when companies and developers build the model into their own systems), Haiku 5.5 costs 0.10 dollars per million tokens for what you send in and 0.50 dollars per million tokens for the response. This applies to requests of up to 100,000 tokens, which according to SiliconANGLE is around 90 percent of them. That is up to 90 percent cheaper than before. For longer requests, the price is 0.50 dollars in and 2.50 dollars out. Anthropic says the model on average costs 75 percent less than Haiku 4.5.

Against competitors, OpenAI's GPT-6 Luna has exactly the same price. On the other hand, SiliconANGLE writes that Haiku 5.5 beats Luna in all six comparable tests. On OSWorld 2.1, which measures computer use, Haiku scores 72.4 percent against Luna's 48.9. On Terminal-Bench 4.0, which measures programming in a command line, the figures are 39.2 versus 16.4 percent. These are Anthropic's own figures, so take them with a grain of salt.

Anthropic has at the same time halved the price of cache reads in Sonnet 5.5, from 20 to 10 cents per million tokens. That can make agent work around 20 percent cheaper.

A monthly price for ordinary users has not been disclosed, and it is not clear whether Haiku 5.5 will come in a free version. It has also not been disclosed whether anything is specific to Denmark or the EU.

What it's good for

Haiku 5.5 is for large volumes of uniform tasks where price and speed count more than deep deliberation. Anthropic mentions summaries, database queries and sorting of inquiries (classification). It is also intended as a helper model for a larger model in programming, and as the engine in customer service and automated browser control.

For a company, this means that AI solutions that were previously too expensive to run at scale can now become a realistic choice. For ordinary users, it is felt mostly indirectly, because apps and services can become faster and cheaper to run.