These 5 Words Cost 12% of Your Claude Code Usage

Hidden Usage from Agent Summaries

Claude Code’s agents panel periodically generates a 3–5-word status label for each active background agent. The video reports that roughly every 30 seconds, Claude sends the agent’s conversation so far to a model to produce the label. Because the conversation grows over time, these calls can use substantial input tokens for very little output.

Measurements

The presenter used Claude Code’s OpenTelemetry export and Better Stack to identify requests tagged as “agent summary.”

• In a test where an agent repeatedly ran sleep 20, summary requests used about 40,000 input tokens each and accounted for 12% of the session’s usage.
• In a more realistic Opus project with three background agents, summaries accounted for about 8%. Those calls also used Opus.
• Flamerix reported 15–20% across team data.
• Anthropic’s figures were lower: about 2% overall and 6% in terminal for the period cited.

The impact depends on how many agents are active and how often their transcripts change.

Why the Cost Is Easy to Miss

The summary calls reportedly do not appear in Claude Code’s JSONL transcripts. As a result, tools that estimate usage from those files, such as ccusage, may omit them.

Prompt caching may reduce the API price of repeated input, but the video notes that Anthropic does not specify how much these calls count against subscription usage.

Ways to Reduce or Avoid the Summaries

1. Set CLAUDE_CODE_FORK_SUBAGENTS to 0. In the presenter’s tests, this stopped agent-summary usage while background agents continued to work. The trade-off is that forked subagents—which inherit the full conversation—are disabled; normal subagents start fresh with only their assigned task.

2. Use Flamerix’s Bun preload workaround. This more technical approach intercepts setTimeout to block the summary timer. The presenter cautions that it hooks into Claude Code processes, so users should inspect the script carefully and pin a reviewed version rather than blindly installing the latest one.

Takeaway

The summaries may help Claude keep track of what agents are doing, but repeatedly processing entire conversations without a simple setting to disable it can consume noticeable usage. The presenter hopes a cheaper implementation Anthropic is reportedly working on will address this.

Main Speakers and Sources

The video features a presenter from Better Stack and discusses findings and a workaround from Flamerix. It also references statements attributed to the Claude Code team at Anthropic, along with user measurements cited from Hacker News and a GitHub issue.