iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
In one developer’s analysis of a month of Claude Code usage, workflow subagents accounted for 48% of their Claude spend, while output tokens made up 0.9% of total tokens. Those are separate measures from one account-specific case study—not typical rates or a general Claude Code benchmark.
What the 48% and 0.9% figures mean
In a September 24, 2026 DEV Community post, author jidonglab reported that workflow subagents represented 48% of their Claude cost. Output tokens accounted for 0.9% of total tokens. The 0.9% figure is a share of tokens, not a share of cost. The post does not specify the calendar start and end dates of the usage month, and it does not provide raw transcripts or an independent audit. Read the author’s account and analysis.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
The author describes a workload weighted toward large audits and research tasks split across agents, and expects subagents to account for less in workflows dominated by single-file edits. The figures should therefore be read as a description of that user’s workload and setup, not as a prediction of your bill.
Why input can outweigh generated output
A subagent’s cost is not limited to the text it returns. In jidonglab’s explanation, an agent makes repeated requests while carrying its initial prompt and context forward alongside conversation and tool results. That can make input usage accumulate over the course of work, even when the final answer is short. The author estimated about 51,000 tokens of starting context per subagent in their own setup, citing system instructions, tool schemas, project instructions, memory, and skill listings as contributors. That estimate is not a documented baseline for all Claude Code installations.
Anthropic’s pricing documentation explains that input, output, cache writes, cache reads, and some tool usage can have distinct cost treatment; tool definitions and tool results also contribute to input. Prompt caching can reduce the cost of repeated input, but it does not make that input free. The documentation does not establish the 51,000-token estimate, prove that every subagent request repeats an identical full context, or verify the author’s cost shares. See Anthropic’s pricing and token-accounting documentation.
What else the author found in their usage
The same analysis reported several other concentrations of spend. These are jidonglab’s own observations and calculations for the workload described in the post, not thresholds that apply to other users:
- 45 sessions costing more than $100 represented 79% of the author’s cost.
- Requests exceeding 400,000 tokens represented 54% of main-session cost.
The author parsed local Claude Code JSONL session transcripts and grouped assistant usage by main thread versus subagent, using the isSidechain field. They summed input_tokens, cache_creation_input_tokens, cache_read_input_tokens, and output_tokens, then applied relevant model rates. Transcript field names and available usage details may vary by version; the author also notes that rates and plans differ. To interpret your own usage, account for the models and routes you used, the token categories recorded in your transcripts, and how the transcript classifies agent messages.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsHow to investigate your own subagent spend
Use the author’s approach as a starting point, not as a ready-made universal calculator. The calculation depends on correctly identifying main-thread and sidechain messages, summing the usage categories present in your files, and applying the rates that match your model and plan.
- Find the session transcripts. Claude Code stores local session transcripts as JSONL files. Identify the files for the period you want to assess, and keep the date range consistent with the billing period you are investigating.
- Separate main-thread and subagent messages. In the author’s analysis,
isSidechaindistinguished the two. Check the field names and structure in your own transcripts rather than assuming they match the author’s version. - Sum token categories by group. The author included input tokens, cache-creation input tokens, cache-read input tokens, and output tokens. Use the categories present in your files and avoid treating cache-related tokens as if they necessarily have the same price as ordinary input.
- Apply the matching rates. Use the rates relevant to the models, routes, and plan involved. Anthropic’s pricing documentation describes the separate treatment of token categories and tool use; do not apply one model’s rate to every session.
- Compare spend and token shares separately. A percentage of cost and a percentage of total tokens answer different questions. Label each metric clearly, as the author’s 48% cost share and 0.9% output-token share illustrate.
Ways to reduce avoidable subagent usage
jidonglab proposed workflow changes aimed at limiting unnecessary delegation and repeated context. The post does not report a controlled before-and-after test or quantify savings from these recommendations; treat them as ideas to evaluate against your own usage.
- Set a default ceiling on agents. Decide how many agents a workflow should use by default, and require a concrete reason to exceed that limit.
- Bundle small tasks. Give one agent a related batch of minor collection or formatting work instead of launching a separate agent for every tiny item.
- Match model effort to the work. Consider a lower-cost model or effort setting for mechanical collection, reserving deeper reasoning for review and synthesis. Anthropic also lists model choice, prompt caching, batching, and usage monitoring among its cost-optimization approaches.
- Trim context that does not help the task. Remove unnecessary global instructions and load project-specific context only when it is relevant.
- Reset stale or oversized sessions. When the topic changes or the session becomes very large, write a concise handoff note and start a fresh session rather than carrying irrelevant history forward.
- Do direct lookups directly. If a lookup is simple enough to answer without delegation, avoid launching an agent just to perform it.
When Claude Code access and billing differ
Claude Code access can depend on the account route. Anthropic’s setup documentation lists Console, Claude plan, and enterprise access routes; the applicable billing and usage details depend on the route and configuration. Check the documentation for the account you actually use rather than assuming that the author’s rate calculation maps to your arrangement. See Anthropic’s Claude Code setup guide.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
