225 billion Claude Code tokens in 30 days, and the subagents that never named a model

说明与用途

作者在30天内使用Claude Code处理2250亿个令牌，发现97.9％为缓存读取。通过显式命名模型、清除上下文和管理代理编排，降低了昂贵的令牌使用并提高了工作流程效率。

问题

High token consumption and inefficient model usage in a large-scale Claude Code agent workflow.

结果

Over 30 days, the author’s agents merged 1,029 pull requests across 21 repositories, produced 1,029 PRs, and reduced expensive token usage by enforcing explicit model naming and context management. Token consumption rose from 4.0 billion to 60.8 billion per week as agent count increased, but workflow efficiency improved with routine task automation.

工作流

• Plan with expensive model, execute with cheaper ones
• Track token usage per agent and per week
• Use /clear between unrelated tasks
• Explicitly name model for each subagent or set CLAUDE_CODE_SUBAGENT_MODEL
• Raise cleanupPeriodDays to preserve logs
• Introduce manager agent to coordinate routine tasks
• Monitor token consumption and adjust agent count
• Switch to cheaper models only after verifying benefits
• Spawn subagents for cheap parts to preserve cache

证据

225.1 billion tokens over 736,633 API responses. Opus models ran 94% of them.

来源

97.90% cache reads, 1.75% cache writes, 0.34% output, 0.01% fresh input

来源

The average response took in about 305,000 tokens of context, 299,000 of them from the cache, and wrote about 1,000.

来源

On Claude Opus 5.5, which did 73% of my tokens, a cache read costs 5% of the input price.

来源

In the week of 31 August my agents used 4.0 billion tokens. Four weeks later they used 60.8 billion in a week.

来源

The work per agent barely changed. Tokens per agent-hour stayed between 52 and 80 million every week.

来源

In October, 97 to 98% of my subagent tokens ran on Opus. In late September and early October, 59 to 80% of my subagents named no model.

来源