In brief: In a blog post, Anthropic outlines practical measures such as /clear, /context, /compact, @-mentions and fixed model selection that let developers specifically reduce token consumption and thus the cost of Claude Code sessions.
In a blog post, Anthropic explains what determines the price of a token in Claude Code and how the number of unnecessarily consumed tokens per session can be reduced. Unlike classic development tools with flat-rate pricing, agent-based coding tools such as Claude Code incur costs for the same task that vary depending on how the work is carried out.
In the post, Anthropic compares two scenarios for the same bug fix: In one session, Claude reads the test and the affected file in a targeted way, makes the change, and is done after a few turns. In another, Claude first searches the entire repository, reading dozens of files on the way to the same two relevant locations – and all of that context read stays in memory for the rest of the conversation, making every subsequent turn more expensive. The result is identical, but token consumption differs significantly. According to Anthropic, efficient token use does not fundamentally mean fewer tokens, but rather that the tokens used actually contribute to the task at hand.
According to the post, the price of a token depends on three factors: the model used, whether it is an input or output token, and whether the token was served from the cache. A larger model does more work on both input and output tokens and is correspondingly more expensive – which model makes sense for which task is addressed by Anthropic in a separate post (“Choosing a Claude model and effort level in Claude Code”). As a rule of thumb: use larger models for genuinely difficult or ambiguous problems, and smaller ones for routine work.
For practical use, Anthropic names several concrete measures: The command /clear should be run between individual tasks so that irrelevant context from previous tasks is not sent to the model again. Model and effort level should be fixed before starting a session, since changing them mid-conversation can invalidate the prompt cache and thereby increase costs. Files should be referenced via @-mention rather than merely named, since this attaches them directly to the message and avoids an additional read or search call.
For commands that produce a lot of output text, Anthropic recommends quiet flags or running them in a subagent, since the command output is embedded into the conversation like a file and remains there for the rest of the session. The command /context should be run once at the start of a fresh session to see what is already loaded (such as CLAUDE.md files or MCP tool definitions) so that unnecessary items can be removed. Before longer breaks, Anthropic recommends /compact, since the prompt cache expires after one hour and summarizing a conversation is cheaper while it is still in the cache.
Source: claude.com · Published August 13, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.