Claude Haiku 5.5 Released — How to Save API Costs with Effort Settings

Anthropic has released Claude Haiku 5.5, which balances speed and cost while maintaining high reasoning capabilities. It features an adaptive thinking function that automatically adjusts the depth of reasoning based on the complexity of the query. Here is a summary of the key changes in Haiku 5.5 and how to practically use it to optimize agent performance while saving on API costs.

What are Adaptive Thinking and the 5-Level Effort Settings?

Previous reasoning models required users to manually specify the token budget for thinking. In contrast, Claude Haiku 5.5 operates by default with an adaptive thinking approach that determines the difficulty of the question and adjusts the length of its reasoning process automatically.

Developers can manually control the depth of reasoning using 5 levels of Effort settings based on the nature of the task. The available options are low, medium, high, xhigh, and max, with the default being medium.

For tasks requiring quick responses, such as simple text conversion or data extraction, it is advantageous to set Effort to low. This reduces unnecessary computation, shortens response times, and saves on token consumption. Conversely, for complex code verification or multi-step agent logic, you can increase it to high or higher to encourage sufficient reasoning.

The 100k Token Tiered Pricing That Determines API Costs

Claude Haiku 5.5's API pricing follows a tiered structure where costs vary based on prompt length. The most important benchmark for cost management is 100,000 tokens.

When the prompt size is 100,000 tokens or less, it is very affordable at $0.10 for input and $0.50 for output per million tokens. However, once the prompt exceeds 100,000 tokens, the cost increases fivefold to $0.50 for input and $2.50 for output per million tokens.

When handling long conversation contexts or large documents, it is important to implement a strategy to keep prompt sizes within 100,000 tokens. If you must continuously pass large-scale context, you can lower costs by actively using prompt caching, which costs only $0.01 per million tokens.

Practical API Code Configuration and Common 400 Errors

Adjusting the reasoning depth of Haiku 5.5 is simple. When using the Anthropic Python SDK, you just need to specify the desired effort level in the output_config object in your API request.