Claude Code’s Hidden Token Drain

Claude Code's background-agent summary feature sends roughly 40,000 input tokens to generate about 130 output tokens every 30 seconds, according to OpenTelemetry trace tests by observability platform Better Stack. The agent_summary requests accounted for 12% of total cost in a synthetic session running multiple sleep tasks and about 8% of token usage in a realistic TanStack build with three background agents, while Anthropic's global telemetry puts the figure at 2% of API cost and 6% in the terminal. The overhead is amplified because summaries default to the primary agent model, often Claude Opus, rather than a cheaper model such as Claude Haiku, and the requests are absent from local .jsonl transcripts, causing transcript-based usage tools to undercount consumption.

The five-word label with a 40,000-token appetite Every 30 seconds, Claude Code polls active background agents , asking a language model to summarize their work in a few words for the agent panel. This seemingly innocuous feature, designed for user visibility, can silently consume a significant portion of a user's token allowance. Better Stack’s OpenTelemetry test, which exported Claude Code’s trace data, revealed the extent of this drain. Summary requests appeared approximately every 30 seconds, each transmitting the agent’s entire conversation history. This context bloat meant requests often involved about 40,000 input tokens to generate a mere 130 output tokens, such as "running seventh 20-second sleep." These summaries, tagged with query source = "agent summary" , accounted for 12% of the total cost in a synthetic session designed to run multiple sleep 20 tasks. In a more realistic scenario—building a TanStack site using three background agents—the summary overhead still consumed about 8% of the session's total token usage . This occurs because the system defaults to using the primary agent model often Claude Claude Opus for these summaries, rather than a cheaper alternative like Claude Claude Haiku. Why a tiny status update keeps getting more expensive Repeatedly, Claude Code sends the entire conversation history —not just recent changes—to the model for each active background agent. This compounds context and input-token volume across polling cycles, growing from 18 messages to 42 in one test. Each summary call consumed approximately 40,000 input tokens to produce a mere 130 output tokens, like "running seventh 20-second sleep." A significant concern is model selection. The video’s tests found these summary calls running on the agent’s assigned model, including expensive Claude Claude Opus https://www.stork.ai/en/claude-opus-4-6 , rather than a smaller, more economical model such as Claude Claude Haiku . This amplifies the cost of a routine UI update. Community measurements of this token drain vary, with agent count and activity driving divergence. Better Stack’s tests found agent summary accounting for 12% of a synthetic sleep session and 8% of a realistic TanStack build. Other community benchmarks ranged from 8% to 26% for heavy multi-agent workflows. Anthropic cited lower averages, with an engineer noting 2% of API cost and 6% in the terminal across global telemetry. Prompt caching may reduce API prices by about a tenth, but Anthropic states subscription usage still draws from quotas without full transparency on the actual cost savings. The usage you won’t find in your usual logs Observing the hidden costs of agent summary requires peering beyond standard logging. Claude Code’s local .jsonl transcripts, often used by tools like cc-usage for cost analysis, do not record these summary requests. Consequently, transcript-based usage calculations can significantly undercount actual token consumption . To verify agent summary activity, enable OpenTelemetry export within Claude Code. Export these traces to an observability platform such as Better Stack. There, filter and group requests by query source = agent summary to reveal their frequency and token volume. This method provides granular, measurable evidence of this specific overhead. Better Stack's tests revealed agent summary accounted for 12% of total session spend in a synthetic sleep test and 8% during a realistic TanStack build. External community findings from Flamerix AI noted up to 16.4% of total spend and 19.7% of API requests attributed to these summaries in high-frequency, multi-agent runs. For more on Claude Code's broader utility, see Claude Code: Best practices for agentic coding - Anthropic https://www.anthropic.com/news/claude-code . Anthropic's telemetry suggests a global average of 2% at API list prices and 6% within the terminal. While these numbers demonstrate meaningful overhead in specific workflows, the exact share varies considerably based on agent activity, task duration, and concurrent processes. It’s crucial to distinguish these measured impacts from universal claims, as individual usage patterns heavily influence the observed token drain. Enjoying this? Get one like it in your inbox each morning. one email a day · unsubscribe in two clicks · no third-party tracking Cut the drain without breaking your workflow Mitigating Claude Code’s token drain doesn't require sacrificing your agent-driven workflow. One effective workaround involves setting the fork subagents option to 0 . This disables subagents that inherit the full conversation history, thereby stopping the summary calls while allowing ordinary background agents to operate with less inherited context. Tests confirm this eliminates agent summary costs. Another solution, albeit more advanced, uses a Bun preload script to hook setTimeout and block the summary timer. This approach modifies Claude Code processes, so scrutinize the code thoroughly. Avoid blindly executing floating curl installs; instead, pin a reviewed, local copy to mitigate security risks. To minimize future drain, actively monitor agent-heavy sessions . Avoid unnecessary subagent sprawl and carefully weigh the convenience of real-time status labels against their cumulative usage cost. Anthropic reports indicate improvements to prompt caching and the potential for cheaper summary options are in development, though a universal fix isn't guaranteed. Frequently Asked Questions What are Claude Code agent summaries? They are short descriptions of what background agents are doing, displayed in the agent panel and generated periodically by a model. How much usage can agent summaries consume? The reported share varies by workload: Better Stack measured 12% in a synthetic test and about 8% in a multi-agent build, while Anthropic cited lower averages across its telemetry. Can I turn off Claude Code agent summaries? There is no straightforward dedicated toggle described in the source. Setting fork subagents to zero can stop these summaries, but changes how forked agents inherit context. Why might usage tools miss these calls? Agent summary requests may not appear in Claude Code’s standard JSONL transcripts. OpenTelemetry data can expose them under the agent summary query source. Are the summary calls always billed at full token prices? Not necessarily. Prompt caching can reduce API costs, but subscription usage accounting is less transparent, so the precise impact depends on plan and workload.