Anthropic Releases Claude Code Session Cost-Saving Guide: Cache Hits Can Reduce Costs by 90%
- Anthropic published six cost-saving recommendations for Claude Code in a blog post summarized by PANews on August 15.
- The guidance says users should clear context with /clear when switching tasks to avoid carrying unnecessary history.
- It recommends deciding the model and reasoning intensity at the start of a task to reduce mid-task switching and preserve prompt caching.
- Anthropic advises attaching code with @file instead of making extra Read calls, and using silent parameters or sub-agents for commands with verbose output.
- The post also suggests using /context to inspect and remove irrelevant loaded content, and /compact to compress a conversation before ending a session.
- Anthropic said output tokens cost about five times as much as input tokens, while cached prompt reads are billed at 0.1 times the normal input price; developers reportedly use about $13 of tokens per day.
Anthropic has published a set of practical cost-control tips for Claude Code, with PANews summarizing the guidance on August 15. The recommendations focus on reducing unnecessary context and avoiding workflow changes that disrupt prompt caching.
The blog suggests using /clear when moving between tasks, choosing the model and reasoning intensity at the start of work, and attaching code with @file rather than issuing additional Read calls. It also recommends silent parameters for verbose commands or routing those commands to a sub-agent.
Anthropic further advises using /context in a new session to review loaded material and remove irrelevant content, and using /compact before ending a session to compress the conversation. According to the post, output tokens are priced at about five times input tokens, while cached prompt reads cost only 0.1 times the normal input price.
Anthropic said the aim is to help developers keep Claude Code usage more efficient, and it estimated daily token spending at around $13 per developer. The report is neutral for markets and mainly reflects product-use optimization rather than a direct financial or trading development.
Anthropic发布Claude Code会话省成本指南：缓存命中可将成本降低90%
Anthropic发布了一组针对Claude Code的实用省成本建议，PANews于8月15日对这份博客内容进行了报道。核心思路是减少无效上下文，并避免在工作流中频繁切换导致提示缓存失效。
这份指南建议在不同任务之间使用 /clear 清理上下文，在开始工作时就确定模型和推理强度，并通过 @file 附加代码，而不是额外发起 Read 调用。对于输出较多的命令，Anthropic还建议使用静默参数，或把任务交给子代理处理。
Anthropic还建议在新会话中使用 /context 检查已加载内容、移除无用信息，并在结束会话前使用 /compact 压缩对话。文章称，输出token的价格约为输入token的5倍，而缓存后的提示读取仅按正常输入价格的0.1倍计费。
Anthropic表示，这些方法旨在帮助开发者更高效地使用Claude Code，并称开发者每天的token支出大约为13美元。该报道对市场的直接指向不强，整体更偏向产品使用优化信息，影响偏中性。