According to CoinMeta, Anthropic has released a cost-saving guide for the Claude model, recommending that developers start with low-intensity inference and only increase it if necessary. This approach can save about 41% of costs while also improving accuracy. The guide suggests that most programming and knowledge-based tasks can begin with a moderate inference intensity of Opus 5.5, while more affordable Sonnet 5.5 can be used for tasks with clear requirements, such as functional development and daily automation. For simple tasks like classification and information extraction, Haiku 5.5 is recommended. For large-scale code migrations and complex scientific research tasks, Fable 5.2 may be considered. Through testing with Swe-bench and Pro programming, it was found that using Opus 5.5 with low inference intensity resulted in a pass rate of about 97%, compared to 95.3% when high inference intensity was used.