Anthropic Publishes the Token Economics of Claude Code: Output Costs ~5x Input, Cache Reads 0.1x, and Subscription Cache Expires in 1 Hour
A 2026-08-14 Anthropic post lays out, for the first time in one place, how Claude Code session choices map to token cost: output tokens run roughly 5x input because decode is sequential, cache reads cost 0.1x input price, and the prompt cache expires after 1 hour on subscriptions versus 5 minutes on API keys. Concrete guidance includes running /clear between tasks, fixing model and effort level at session start to avoid busting the cache mid-conversation, @-mentioning files to skip a Read call, adding quiet flags or delegating noisy commands to subagents, and running /context in a fresh session to find tool definitions you are paying for and not using. For anyone running agents on a subscription rather than the API, the 1-hour cache TTL is the single number that should reshape how sessions are paced.
↳ Follow the thread