TokenSave Data · Dataset

What the Claude Code harness costs: plugins, subagents, plan mode, skills and hooks

Most of what Claude Code sends to the model is its own harness, not your task. The same small coding task used a median 78,723 input tokens in a default setup and 8,214 with a one-line system prompt and two tools (−90%); every game still worked. Delegating the task to a subagent cost 128% more, and installed plugins added cost even when none was used.

Measured 2026-10-19 · Claude Code (Opus 5.5) · 7 settings · 5 runs each · License CC BY 4.0 · Related: Claude Code vs Codex: startup tokens

DOI 10.5281/zenodo.23465633 · Also on Hugging Face · Kaggle · GitHub

Key findings

• The harness is about 90% of the input. With Claude Code’s system prompt and tool descriptions replaced by one sentence and two tools (Write and Read), the median run used 8,214 input tokens instead of 78,723 (−90%) and cost $0.13 instead of $0.37 (−64%). All 5 games worked. The games were also shorter (median 7,904 characters against 13,375), so part of the saving is output.
• Installed plugins cost tokens even when unused. This laptop has 35 plugins installed, which list 376 skills and 19 agents to the model. --safe-mode loaded 3 plugins, 19 skills and 5 agents, with the same 27 tools. No run used a plugin, yet safe mode cut the median input by 39% (30,660 tokens per task) and the cost by 54%. Here too the games came out shorter (7,921 characters, 4,718 output tokens against 7,364), so not all of the cost drop is input.
• A subagent is the most expensive way to do a small task. Asked to delegate the work with the Agent tool, Claude Code did so in all 5 runs. The median run used 534,125 input tokens (+578%), cost $0.85 (+128%) and took 178 seconds instead of 61 (+192%). The subagent starts with its own copy of the harness and checked its work with several shell commands.
• Plan mode roughly tripled the input but added a quarter to the cost. Planning first and then building (two calls, the second continuing the session) used 245,562 input tokens (+212%) and cost $0.47 (+26%). Most of the extra input was read back from the prompt cache, which is billed at a fraction of the normal price.
• A skill or a hook doubles the token count but barely moves the bill. One project skill (a 1.9 KB SKILL.md, used in all 5 runs) raised input to 158,393 (+101%) but cost only +2%; a PostToolUse hook that sends a reminder after each file write raised input +54% and cost +3%. Each adds model turns, and every turn re-reads the whole context, but nearly all of it from the cache.
• Every game worked. 35 of 35 games loaded and responded to keyboard and mouse in a headless browser without script errors, so none of these settings bought or lost correctness on a task this small.

Tokens, cost and time by setting

Median of 5 runs per setting. Input tokens include cache reads and cache writes. Cost is the API list-price estimate Claude Code reports; subscription plans are not billed per token. Every run is in the CSV.

What this means for our other Claude Code datasets

Our earlier Claude Code measurements, such as the 35,771 input tokens for a one-word reply, were made on this same laptop with the same 35 plugins installed. Comparisons inside each dataset (file sizes, languages, effort levels) are unaffected, but the absolute Claude Code numbers include the cost of those plugins, so a clean installation will start lower.

The settings

• Default: claude -p --effort medium --output-format stream-json --verbose --permission-mode acceptEdits with an empty MCP config (--strict-mcp-config). The task prompt: “Make a brick-breaker game that runs in the browser as a single file named index.html in the current folder. It needs 2 levels, a score display and 3 lives, and it must be playable with the keyboard and the mouse. Write the file and finish.”
• Safe mode: the same plus --safe-mode, which leaves out personal plugins, skills, hooks and CLAUDE.md.
• Minimal harness: --system-prompt "You are a coding agent. Use your tools to complete the user's task." --tools Write,Read.
• Subagent: the default command with one more sentence in the prompt: “Delegate the implementation to a subagent with the Agent tool.”
• Plan mode: the prompt with --permission-mode plan, then claude -c in the same folder with “Go ahead and implement your plan.”. The tokens and cost of both calls are added together.
• Skill: a 1.9 KB project skill, .claude/skills/html-canvas-games/SKILL.md, with conventions for single-file canvas games, placed in the working folder.
• Hook: a PostToolUse hook on Write and Edit that returns a reminder to check the requirements before finishing.

Method

• Claude Code 2.1.295 with Opus 5.5 on one Windows laptop, effort medium, no MCP servers, each run in its own empty folder, the prompt sent through standard input. The seven settings were run in turn, five rounds, on 2026-10-19.
• Tokens and cost come from the modelUsage and total_cost_usd fields of the final result event; the numbers of tools, skills, plugins and agents from the system/init event.
• “Game works”: the file loaded in headless Chromium with a canvas, took mouse clicks and arrow keys for about two seconds and threw no script errors.

Limitations

• One small task and five runs per setting; on larger tasks a skill, a plan or a subagent may save rework that this task never needed.
• The plugin overhead depends on what is installed; 35 plugins is a heavy setup, and fewer plugins will cost less.
• The minimal harness has no shell, search or edit tools, so it is a measure of overhead, not a recommended setup.

How to cite

Free to use, quote and republish with attribution under CC BY 4.0. Please link to this page.

Embed the chart

Related

• Claude Code vs Codex: startup tokens and the same coding task
• CLAUDE.md and AGENTS.md size
• Claude effort levels: tokens, time and cost

Questions or corrections: contact@jonhisking.com