Anthropic explains how Claude Code burns tokens

Anthropic's post draws a distinction that matters for anyone paying per token: subscription-style coding tools hide the cost of a request, but agentic tools like Claude Code price each task separately. The post walks through where those tokens go, splitting the work into a prefill phase and a decode phase. In prefill, the model reads the request, the context, the system prompt, CLAUDE.md, the message, and everything added to the conversation since the last message, including files Claude read and command output. In decode, output tokens — the thinking, tool calls, and text — are generated one at a time.

Much of the guidance circles back to the prompt cache, which Anthropic describes as the mechanism that keeps costs down. A cached read costs 0.1 times the input price, and a write to the cache costs up to 2 times the input price but happens only once per token. Anthropic warns that user actions can invalidate the cache anyway: caches require a match at the start of the request, so switching models or effort levels mid-session, switching permission modes, running /compact, or simply leaving time to pass can force the whole conversation to be re-read at full price. The one-hour expiry after last use and the five-minute expiry after the API key's last use are the main culprits.

Anthropic's summary names four factors to watch as having the biggest effect on cost: long sessions, carrying too much context, using a larger model or higher effort level than the task needs, and a broken prompt cache. The practical suggestions — running /clear between unrelated tasks, setting /model and /effort before starting, attaching files rather than asking Claude to read them, checking /context early, and running /compact before the cache expires — all push in the same direction. Whether they bring bills down in practice is likely to depend on how carefully individual engineers follow the session-start advice, since the cheapest path Anthropic describes (cache reads at 0.1 times input) is also the easiest to lose to a mid-session change.