Claude Haiku 5.5: cheaper small model with 1 million tokens

What's new

On October 7, 2026, Anthropic launched Claude Haiku 5.5, the latest version of the company's smallest and fastest model family. Anthropic calls it the cheapest, fastest and most capable small model the company has released, and says it is the company's fastest model to date at standard speed.

The model can hold far more in memory at once. The context window, meaning how much text it can keep track of in a conversation or task, has grown from 200,000 to 1 million tokens. A token is a small piece of text, often part of a word. The maximum length of a response has doubled from 64,000 to 128,000 tokens. The model's knowledge cutoff is June 2026.

What's smart

Haiku 5.5 is the first Haiku model with an adjustable effort setting. Users can choose between Low, Med, High, Xhigh and Max, and so decide for themselves whether they want the cheapest and fastest answer or want the model to put more effort into thinking things through.

On Anthropic's own benchmarks (standard tests used to compare models), Haiku 5.5 scores 72.4 percent on OSWorld 2.1, where the AI controls a computer, 39.2 percent on Terminal-Bench 4.0 and 1,620 Elo on GDPval-AA v2.1. On Humanity's Last Exam, a very difficult knowledge test, it scores 45.9 percent without tools and 57.4 percent with them. According to a summary of Anthropic's figures, its predecessor scored 15.7 percent on OSWorld 2.1 and 735 Elo on GDPval-AA. The figures come from the manufacturer and have not been independently verified.

Cheaper or better?

Both, but the price drop is not quite as large as the headlines may suggest. The API price (for companies and developers building on the model) is $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. For longer prompts, the price is $0.50 for input and $2.50 for output. Haiku 4.5 cost a flat $1 for input and $5 for output.

Anthropic says the price is 90 percent lower for requests up to 100,000 tokens. A footnote reveals, however, that the new tokenizer, which splits text into tokens, produces about 30 percent more tokens for the same text. As a result, the average saving is about 75 percent. For the longest requests, the saving is about 50 percent. According to the footnote, 90 percent of requests to Haiku 4.5 were under 100,000 tokens.

Caching, where the model reuses text it has already seen, costs $0.01 per million tokens to read up to 100,000 tokens and $0.05 above that. Anthropic has also halved cache reads for Sonnet 5.5 to $0.10, which the company says results in about 20 percent lower total cost.

Customers report improvements: Asana says latency is more than 30 percent lower and responses are up to 2.5 times faster. Box measures a score 11 points higher than with Haiku 4.5 at about half the latency. AlphaSense gets 0.84 versus 0.76 on questions about documents. The monthly price for ordinary users and a comparison with competitors' models are not given in the sources.

What it's good for

The model is built for tasks where speed and price matter: customer service, processing large volumes of documents and automation in which an AI needs to work with the computer. The large context window makes it possible to give it long reports or many files at once.

Haiku 5.5 is available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry/Azure under the name claude-haiku-5-5. The sources do not say whether separate prices or terms apply in Denmark and the EU.