Claude Code background summaries can account for a large share of total costs

The hidden cost of status updates

Claude Code includes a feature that allows agents to run in the background while a user continues other work. To keep the user informed, the tool provides a progress panel that displays a short 3-to-5 word description of what each agent is doing at any given moment.

As reported by theainewsreport.com, this status line is not generated by the agent itself. Instead, Claude Code forks a request every 30 seconds, sending the agent's entire conversation transcript to a model—usually Opus—to summarize the most recent action.

Because these requests happen so frequently, they can add up quickly. One team measured these summary calls as making up 16.4% of their weekly spend, while other measurements on GitHub suggest the cost could range between 18% to 26% of total requests.

How the billing math works

The cost of these calls is heavily influenced by the size of the transcript and the effectiveness of the prompt cache. A prompt cache allows the model to reuse previously processed data, making subsequent reads much cheaper than the initial input.

However, even with a cache hit, the cost is not zero because the model is billed for every token it re-reads. For a 200,000-token transcript, a cache hit costs about 4 cents, but a cache miss can cost between 80 cents to $1.60 depending on the specific cache write type.

The data suggests that while misses are rare—occurring in only about 2% of calls—they are responsible for a large portion of the bill. Because a miss requires re-sending the entire transcript at full price, those few instances can account for over a third of the total summary costs.

Who is affected by these costs

Not every user will see a significant impact on their bill. The system includes skip rules that prevent charges when an agent is idle, waiting for a permission prompt, or when the transcript has not changed since the last summary was generated.

The costs primarily target users who run active background agents or forked subagents that work for several minutes at a time. The more agents a user runs simultaneously, and the longer their transcripts grow, the higher the cumulative cost of these 30-second updates will be.

Notably, these requests do not appear in the session logs stored on disk. This means that standard tools used to monitor and sum transcript costs may undercount the actual spending, as the summary calls are handled as separate background queries.

Current limitations and workarounds

Currently, there is no official switch in the Claude Code settings to turn off these specific progress summaries. While the documentation mentions a progress summary option, it only serves to turn the feature on, and setting it to false does not change the behavior.

Users who want to disable the feature can use environment variables, but these only work if they also disable background agents entirely. A third-party preload script exists for Linux users to block these calls, but it is not officially supported and may fail if the software is updated.

The most direct way to address this is through the official GitHub repository. A community issue has been opened requesting a supported way to disable these summaries, which would allow users to keep background agents active without the recurring cost of the status updates.

Measuring your own usage

Because usage patterns vary wildly between individual developers and large teams, users should measure their own costs before making changes. Claude Code provides OpenTelemetry exports that can be used to track specific request types and their associated costs.

By exporting these logs, users can group requests by the query_source attribute. Looking specifically for the agent_summary tag will reveal exactly what percentage of the total cost_usd is being spent on these status updates.

If the summary share is high, users can optimize by running fewer long-lived background agents or assigning cheaper models to subagents. Sharing these findings on the official GitHub issue can also help Anthropic see how these costs impact different types of workflows.