Claude Code Calls the Model Every 30 Seconds Just to Report Progress
An outside analysis found that Claude Code ships the conversation history of background work to the model roughly every 30 seconds. The share of total cost attributable to those requests varied widely depending on the measurement setup, and someone identifying themselves as a Claude Code developer said one expensive case has already been fixed.
Flamerix AI has published an analysis of how Claude Code displays the progress of ongoing work, along with unofficial code to block those requests. According to the write-up, Claude Code sends the conversation history of an AI working in the background to the model roughly every 30 seconds in order to generate a short progress message.
Summaries every 30 seconds: how they are produced
The subject of the analysis is the AI that keeps working behind the scenes, known as an agent. An agent is an AI that handles tasks on its own without a person watching over it. To show how far that work has gotten, Claude Code sends the model a separate request for a summary. In other words, on top of doing the job it was handed, it also calls the model to report on its progress.
This behavior is tied to how subagents run by default. Anthropic’s public records show that, as of 2.1.232, non-team-member agents spawned from an interactive session run in the background by default. The range Flamerix AI examined in this write-up is Claude Code 2.1.257 through 2.1.312 on Linux.
Someone identifying themselves as a Claude Code developer also explained that an active background agent generates a summary request roughly every 30 seconds. They added that agents sitting idle, with no work underway, do not receive these requests.
How much do the summaries cost?
Flamerix AI said that over the period they measured, status summary requests accounted for 16.4% of estimated cost. By request count, the share was 19.7%.
The figures the developer cited across all users were far lower. Among users who send usage records to Anthropic, the cost of summary requests the previous week came to about 2% of total cost when converted to list prices. The share by request count was under 4%. The developer suggested that Flamerix AI’s numbers may reflect an environment that leans unusually heavily on background agents.
The two numbers cover different populations. One is an estimated cost in a specific setup; the other is list-price-equivalent cost across all users who share usage records.
What has already been fixed
The developer said that as of 2.1.311, one case that was driving costs sharply higher has been fixed. The problem was that some summary requests could not reuse previously processed content, which inflated the cost. They also said work is underway on an approach that sends less content to a smaller model for summarization. If you want to check your own costs, start by looking at which version you are running.
Anthropic includes South Korea among the regions where Claude is available. For users in Korea, the takeaway here comes down to a single point: model usage comes not only from carrying out the work you assign but also from showing how far that work has come.