The September cut took 17% of my Claude Code week. Subagents were taking 48%.

A developer analyzed their Claude Code usage logs to understand the impact of recent usage limit adjustments. By auditing 455 sessions and 63,000 requests, the author discovered that subagents were responsible for 48% of their total token consumption, largely due to high overhead costs during initialization. The analysis highlights that while common optimization tips—such as removing MCP servers or filtering shell output—have a negligible impact, significant savings can be achieved by pinning smaller models to subagents and managing prompt cache TTL settings. The author also notes that long idle periods between sessions lead to massive cache rewrites, suggesting that clearing sessions or using summaries can help preserve usage limits. Ultimately, the study emphasizes that individual workflows drastically change how AI usage caps are experienced, and users should audit their own logs to identify the primary drivers of their token consumption.