Claude settings guide

Put standing context in a Project: upload reference files and write project instructions once, so you stop pasting the same background into every chat.

Match the model to the job: Haiku 4.5 for quick, cheap tasks, Sonnet 5.5 for everyday work and coding, Opus 5.5 or Fable 5.1 only for the hardest multi-step problems. On the API, Haiku costs $1/$5 per million tokens against $10/$50 for Fable.

Ask for the format you want up front ("Answer in 5 bullets, no preamble"). Shorter outputs cost fewer tokens and are easier to check.

Structure long prompts with XML-style tags such as <context>, <task> and <format>. Anthropic recommends this in its prompting docs because it keeps instructions separate from material.

Turn extended thinking on for analysis, maths and planning, and off for simple questions. Thinking tokens are billed as output.

On the API, always set max_tokens and use prompt caching for long, repeated system prompts; cached input is billed at a fraction of the normal rate.

Start a new chat when the topic changes. Long threads resend history with every message, which costs more and can blur the answer.