Anthropic's Cost Playbook for Claude Code: Output Tokens Run 5x Input, and Context Accumulation Is the Real Budget Leak

Anthropic published a practical cost-reduction guide for Claude Code sessions, authored by Lydia Hallie, built around the point that the same coding task can cost wildly different amounts depending on how the session is run. It identifies four cost drivers — model choice, the input-to-output token mix, prompt-caching mechanics, and context accumulation over a session — with concrete multipliers: output tokens cost roughly 5x the price of input tokens, cache reads run about 0.1x standard input pricing, and cache writes can cost up to 2x normal input pricing. Recommended habits: run /clear between unrelated tasks, fix the model and effort level at the start of a session to preserve the prompt cache, use @-mentions instead of triggering redundant reads, run /context at the start of a session to audit what’s loaded, and run /compact before breaks to make the most of already-cached history. Anthropic’s framing: “being efficient with tokens doesn’t mean using fewer of them overall. It means making sure the ones you do use go towards the thing you actually asked for.”

The guidance lands three weeks after Anthropic’s own case for reducing prompt bloat in the first place: a July post on “context engineering” for Claude 5-generation models described rewriting Claude Code’s system prompt to cut over 80% of its instructions without hurting coding-eval performance, on the theory that rigid rules were compensating for older, weaker models rather than helping current ones. That release introduced a /doctor command specifically to let teams audit and rightsize their own context setups — the same audit habit this new guide recommends running at the start of every session.

Together the two posts describe a shift in how Anthropic wants teams thinking about Claude Code spend: not a lever to pull once (cut the system prompt) but session hygiene practiced continuously (clear, cache, audit, compact). For teams running Claude Code at scale, the multipliers matter more than they look — a session twice as long doesn’t just cost twice as much once uncached context and output-heavy responses stack up, which is exactly the creeping spend /context and /doctor are built to catch before it shows up on a bill.