Claude Haiku 5.5 arrives with lower cost and higher speed

Anthropic has introduced Claude Haiku 5.5, an AI model designed to respond quickly, process large volumes of tasks and reduce the cost of running agents. The company positions it for tasks such as summarizing documents, classifying requests, automating browser tasks and assisting with straightforward programming.

The difference lies in its position within the Claude family. Haiku is not designed to solve the most complex problems. Instead, it handles many small, well-defined tasks while larger models take care of planning or deep reasoning.

A model built to work at scale

Claude Haiku 5.5 is available to Free, Pro, Max, Team and Enterprise users on Claude.ai, both on the web and on iOS and Android. It can also be used through the Claude platform, Amazon Web Services, Google Cloud, Microsoft Foundry and Claude Code.

In the API, the price for requests of up to 100,000 tokens, a unit that measures the amount of text processed, is $0.10 per million input tokens and $0.50 per million output tokens. For longer requests, the cost rises to $0.50 and $2.50, respectively.

Anthropic also offers up to 90% savings through prompt caching, which avoids processing repeated instructions again, and 50% savings through batch processing.

What it can do

The company outlines several concrete uses for Haiku 5.5:

• Summarize large amounts of text and classify documents or requests.
• Handle chats, voice calls and support queries in real time.
• Act as a subagent, carrying out specific tasks within a workflow directed by a more powerful model.
• Fill out forms, enter data and move information between applications.
• Make specific code changes and use tools across multiple steps.

The model includes effort controls for the first time. This lets you adjust how much reasoning capacity it uses for each task and find a balance between quality, speed and price.

Results reported by Anthropic and its customers

Anthropic says Haiku 5.5 outperforms Haiku 4.5 in programming, tool use, computer automation and agent tasks. The company also shares customer results, although these tests were conducted in its own products and scenarios.

In Box AI, for example, it scored 11 points higher than Haiku 4.5 with approximately half the latency in initial evaluations. HubSpot recorded an average score of 92.8% across three tests involving CRM tasks, such as reviewing opportunities and detecting old or ambiguous records.

Ask in Document, a company that processes around 8 million calls per week, achieved a statistically significant improvement across 400 queries: 0.84 compared with 0.76 for Haiku 4.5. In another case, an AI agent product recorded a reduction of more than 30% in latency and up to 2.5 times more speed per agent turn.

These figures do not mean Haiku 5.5 is always the best model. They describe results from specific tests and do not replace an evaluation using each company’s own data and tasks.

What changes for you

If you use an application that summarizes emails, answers queries, reviews documents or automates repetitive steps, a model like Haiku 5.5 could make those features respond faster and cost less to operate. For the end user, the most noticeable effect will be less waiting.

For developers, the lower cost makes it possible to run more tasks in parallel. A large model can decide what to do, while Haiku 5.5 searches for a figure in a report, classifies results or completes a specific change.

Anthropic’s page also mentions Claude Haiku 4.5 as an earlier generation. The main development described in the availability and pricing information is Haiku 5.5, which occupies the role of the small model focused on speed and volume in the current offering.

The key question will be how it performs outside each provider’s demonstrations and evaluations. If it maintains its accuracy on repetitive tasks, Haiku 5.5 could become the low-cost component that lets many products add AI without sending every request to a more expensive model.