# App install token benchmark (2026-09-30)

Claude Code 2.1.285, `--model sonnet` (claude-sonnet-5-5), `claude -p --output-format json --max-turns 60|80`,
run on `shawn` with Zeltro installed. Each app was installed twice by the same agent: once with plain Docker
("agent alone", its own directory and host port) and once told that Zeltro is installed ("agent + zeltro").
Prompts are in the two run scripts. Costs are Claude Code's own `total_cost_usd` at Anthropic list prices.
Both runs of each pair succeeded; everything was removed afterwards.

Installing the same app from Zeltro's app library (GUI) or with `zeltro install <app>` uses no AI at all.

Used by resources/pages/ai-app-builder-cost-comparison.md.