Claude Haiku 5.5: 1 million tokens of memory and low output pricing

What's new

Anthropic has added a new model to its Claude 5.5 series: Claude Haiku 5.5. Its model ID is claude-haiku-5-5, and there is also a variant with thinking, in which the model spends extra time considering its answer.

The model has a context window of 1 million tokens. Tokens are the small chunks of text AI models read and write in, and the context window is the amount of text the model can handle at once. The maximum response can run up to 128,000 tokens.

What's smart about it

The large context window means you can give the model a lot of material in one go, such as long reports or large collections of documents. And with output of up to 128,000 tokens, it can also write very long responses.

The model can be used in several ways. According to APIYI, a reseller of AI access, it can be reached through OpenAI-compatible endpoints, through Anthropic's native protocol and through an enterprise channel at APIYI. That makes it easier to plug it into systems that are already built for other models.

For those upgrading from Haiku 4.5, there is a practical detail: The temperature, top_p and top_k settings, as well as assistant prefill, must be removed. Otherwise you risk a 400 error. It is a small thing for a developer, but worth knowing before you switch.

Cheaper or better?

APIYI lists an output price of $0.50 per million tokens as long as the request is up to 100,000 tokens. If it is longer, the price is 5 times as high, or $2.50 per million tokens. According to APIYI, this matches the official price list. The input price is not clearly stated, so it is not mentioned here.

Anthropic says the model is on average about 75 percent cheaper to run than Haiku 4.5. Some posts indicate that 75 percent is a blended average (about 90 percent lower under 100,000 tokens and about 50 percent lower above that), but this is not confirmed. The figure comes from the maker itself and has not been measured by independent parties.

The sources do not say whether or when the model is available in Denmark or the EU, or what it costs ordinary users per month. There is also no comparison with competitors' models, and there are no independent measurements of speed or quality. Whether it is better than its predecessor therefore cannot be determined from the sources.

What it is good for

On paper, the model suits tasks involving a lot of text: reviewing long documents, summarizing large volumes of material or writing long texts. The low output price makes it interesting for companies that send many requests through an AI model.

It is first and foremost a tool for developers and companies that build AI into their own solutions. Ordinary users will typically encounter the model indirectly, through services that use it.