Anthropic Launches Claude Haiku 5.5 With Lower Costs and Better Performance

Anthropic has launched Claude Haiku 5.5, a new artificial intelligence (AI) model built for businesses and developers that need to process large amounts of information efficiently. It is designed for routine tasks such as summarising documents, searching databases, organising information and handling data at scale.

The new model focuses on making AI more affordable for frequent use while improving its ability to complete practical tasks. These changes could benefit companies that rely on AI agents to automate repetitive work and manage complex workflows.

One of the biggest changes is the reduction in operating costs. Anthropic says Claude Haiku 5.5 can cost around 75% less to run on average than its predecessor, Claude Haiku 4.5. For prompts containing up to 100,000 tokens, the model is priced at $0.10 per million input tokens and $0.50 per million output tokens. Longer prompts are also cheaper than before, making the model a more affordable option for businesses that process large volumes of information.

Anthropic has also reduced the cost of cache reads for Claude Sonnet 5.5 by 50%, bringing the price down to $0.10 per million tokens. According to the company, this change could lower the cost of running many AI agent tasks by around 20%.
These pricing changes are intended to make it easier for businesses to use AI for repetitive, high-volume tasks without incurring excessive costs.

Claude Haiku 5.5 also brings performance improvements compared with the previous generation. In Anthropic's published results, the model scored 72.4% on the OSWorld 2.1 offline test, compared with 15.7% for Claude Haiku 4.5. The model has also been evaluated on tasks involving problem-solving and AI-assisted coding. These capabilities are important for developers and businesses looking to automate computer-based work, process information and build AI-powered applications.

However, benchmark results reflect performance on specific tests and do not necessarily guarantee the same improvements across every real-world task. Another new feature is an adjustable effort setting, which allows developers to decide how much effort the model should spend on a task. For simpler requests, developers can prioritise lower costs and faster responses. For more demanding tasks, they can allow the model to spend additional effort to produce stronger results.

This flexibility could help businesses balance performance, response time and operating expenses based on their needs. Early customer feedback has also been encouraging. Software company Asana reported more than a 30% reduction in task completion delays and processing speeds of up to 2.5 times faster for each AI agent turn compared with its existing model. These results come from Asana's own testing and may vary depending on the application and workload.

Claude Haiku 5.5 is available through several platforms, giving developers different ways to integrate it into their applications. The model can be accessed through Anthropic's Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure. Developers can connect to it using the API model ID claude-haiku-5-5. Anthropic is also introducing monthly API credits for eligible Claude Max and Team subscribers. In addition, its Python and TypeScript software development kits (SDKs) are receiving beta support for computer and browser use, opening up more options for building AI-powered workflows.

Claude Haiku 5.5 reflects Anthropic's effort to make AI automation more practical for organisations that need to complete large numbers of routine tasks. Lower prices could make it easier to scale AI use, while improved benchmark performance and adjustable effort settings give developers more flexibility.

The main advantage is the combination of lower costs, task-focused performance and greater control over how the model works. For businesses using AI agents or processing large volumes of data, these changes could help make automation more affordable, although actual savings and performance will depend on the workload and implementation.

This article is based on information from Techgig