Anthropic has published a technical breakdown of Claude Code token usage that says developers spend about $13 a day on average, or roughly $150 to $250 a month, and that the same job can cost several times more depending on how a session is run. The company says every turn re-sends the full prior context, so files read, command output and earlier dialogue make each later turn more expensive unless the prompt cache stays intact. It frames costs around two stages, prefill for reading all input context and decode for generating replies and reasoning tokens, with current list prices of $5 per million input tokens and $25 per million output tokens for Opus 5, $2 and $10 for Sonnet 5, and $1 and $5 for Haiku 4.5. Anthropic's six main recommendations are to clear sessions after each task, keep the same model and reasoning effort from the start, attach files with @ references, suppress noisy command output, run /compact before cache expiry and offload large-output work to subagents. It also reiterates earlier advice to use /rewind for recent missteps, keep prompts specific, inspect startup context with /context, favor smaller models for routine work and run /loop in a separate clean session. The company says prompt cache hits can cut repeated input costs to one-tenth of the normal rate, but the cache breaks if users switch models or effort, toggle Fast mode, compact the conversation, let the cache expire or resume an old session, and it flags opusplan as an added trap because entering and exiting the plan phase switches models. Anthropic says it now writes 80% of its code with AI, has increased code merge volume 8x in a year and accelerated benchmarks 52x, making token discipline a practical engineering and budget issue.