Claude Code Agent Summaries Ate 16.4% of One Team’s Token Spend

Every 30 seconds, for every background agent it runs, Claude Code sends that agent’s full transcript to the agent’s own model and asks for a 3-5 word description of what it just did. One team says those few words cost more than the actual work. Over one week, the summaries took 16.4% of its estimated Claude Code spend, according to Flamerix AI’s write-up, against 15.7% for the main conversation.

The text only feeds the per-agent status in Claude Code’s agents panel.

What happened

Flamerix builds a monitoring tool called DRIM, and it spotted the pattern in Claude Code’s OpenTelemetry export. Grouping API requests by their query_source field put one label, agent_summary, at 19.7% of all requests. About 96% of that spend ran on Opus, at roughly 9,000 input tokens for every output token.

The post went up at 03:33 UTC on Thursday and reached the Hacker News front page within hours. Its headline blamed Claude Code’s status line, which caused some confusion. Anthropic’s status line docs say that feature runs locally and uses no API tokens. The cost sits in the agent summaries, not the bar at the bottom of your terminal.

Flamerix isn’t the only one counting. GitHub issue #86879 logged 339 summary calls in a single four-hour session with 27 subagents, or 22.7% of all requests. A second user on the same thread found the summaries ate 26% of cache-read tokens in a session that lasted under two hours.

What this means for freelancers and students

If you pay for the AI tools freelancers rely on out of your own pocket, this lands on your budget, not a company’s. On API billing it shows up as dollars. On a subscription it shows up as lost rate-limit headroom, which means hitting your cap sooner during client work or a deadline week.

The cost grows with how you work. More concurrent agents and longer contexts mean bigger summaries, more often. One commenter on #86879 measured an agent with 380,000 tokens of context paying about 45 million cache-read tokens an hour just for its spinner text.

Most of these calls hit the prompt cache, which makes them cheaper but not free. Flamerix says roughly 98% read from cache, and the remaining 2% carried over a third of the summaries’ cost.

If you mostly run one plain session without background agents, you may not be paying this at all. Nobody has confirmed that yet.

The spend your logs don’t show

The bigger problem is visibility. Popular cost trackers read the session transcripts Claude Code writes to disk, and these requests aren’t in them. One open GitHub issue, filed October 22, found transcript-based totals running 1.35 to 2.13 times lower than Claude Code’s own cost tally across 14 sessions. So a freelancer checking a weekly spend report could be looking at a number that’s quietly wrong.

Anthropic’s cost documentation says background processes typically use under $0.04 per session. Its list of background usage covers conversation summarization, some commands and prompt suggestions. Agent summaries aren’t on it.

There’s no clean switch, either. “We couldn’t find a supported setting that disables only these summaries,” the Hacker News poster wrote. Two environment variables, CLAUDE_CODE_FORK_SUBAGENT=0 and CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1, do stop them. They also switch off fork subagents or background agents, which is the workflow most heavy users want to keep.

Flamerix’s own fix is an unofficial Bun preload that blocks the summary timer without modifying the Claude Code binary. In its tests across versions 2.1.257 to 2.1.312, the same agent run made five summary calls without the hook and zero with it. Flamerix has only tested it on Linux, and macOS and Windows users are on their own for now.

What to watch next

No Anthropic response had appeared on either GitHub issue or the Hacker News thread by Thursday morning. The company also hasn’t said how plan limits weigh cached reads, which decides how much this matters to subscribers.

It does have one option already built. Version 2.1.310 ships a cheaper short summary mode that uses about 2,000 tokens, but it’s switched off.

The summaries have been on by default since version 2.1.232, published August 13. Watch whether Anthropic adds a setting, turns that short mode on, or starts writing these requests to disk where cost tools can see them.