A Hidden Status Label in Claude Code Reportedly Eats Up to a Quarter of Usage
A developer's telemetry suggests Claude Code's background agents re-send their whole transcript every 30 seconds just to write a 3 to 5 word status line. Here is what was measured and what is still unverified.
A write-up that climbed to the top of Hacker News overnight claims Claude Code spends a surprising share of its budget on something you never asked for: a tiny status label. According to the author, a feature called “agent_summary” accounted for 16.4 percent of their Claude Code spend, slightly more than the main conversation itself (15.7 percent). Other users in a linked GitHub issue reportedly measured between 18 and 26 percent. These are one person’s numbers, not an Anthropic statement.
Here is the mechanism as described. Claude Code can run background agents, small helpers that work on a task while you do something else. About every 30 seconds, for each of them, the tool sends the agent’s entire transcript to the model and asks it to “describe your most recent action in 3-5 words.” The answer only updates the status line in the agents panel. It never reaches the conversation. The author counted roughly 9,000 input tokens (the chunks of text a model reads, billed per piece) for every output token. Most of that spend, 96 percent, reportedly landed on the Opus model. The author says the feature has been on by default since version 2.1.232 in August and that no setting switches it off.
The author found it by grouping their usage logs by request type, which showed this one category costing more than expected. Their workaround is a small script that intercepts the timer behind the feature. After applying it, the agent_summary cost dropped to zero in their logs, and the only visible loss is that the agent panel stops updating its status.
What’s behind it: every time a model is asked something, it has to re-read everything before it. A long transcript is cheap to store but expensive to re-read again and again. A feature that looks tiny on screen can therefore be the biggest line on the bill, simply because it runs often on a long history. This is a common trap in agent tools, and it is why usage breakdowns matter.
Some honest caveats. The dollar figures are Claude Code’s own per-request estimates, not invoices. The workaround matches text inside the program, so it could break silently after an update, and it was tested only on Linux with versions 2.1.257 to 2.1.312. The author also notes that altering a client’s behaviour may conflict with the service’s terms. Anthropic has not commented in the material we reviewed.
What this means for you: if you use Claude Code casually, there is nothing to do today. If you run several background agents and keep hitting rate limits on a subscription, it is worth checking your own usage breakdown before changing anything. I would wait for an official setting or fix rather than patching the program yourself.
Sources
Source: https://gist.github.com/flamerix-ai/c3905705cc4205e5de3011fa2e74efce
Claude for Legal Now Checks Its Own Sources Before a Lawyer Looks
Anthropic added interactive overviews and 60 plus integrations to Claude for Legal, with automatic source checks that hand the file to a human lawyer.
