Maximizing the value of your Claude Code sessions | Claude by Anthropic
Claude Code charges per token used in a session, with input tokens costing less than output tokens and cached tokens costing far less than fresh ones—meaning the same task can cost different amounts depending how you run it. The post explains that token costs multiply based on model size and effort level, then get amplified by prompt caching behavior: every turn resends the entire conversation history (cheaply, from cache) plus whatever is new, so keeping conversations focused and clearing between unrelated tasks saves money by preventing irrelevant context from accumulating. Practical tactics like @-mentioning files, using /clear and /compact strategically, and suppressing verbose command output all reduce what stays in the context window, which compounds savings across multiple turns.