Claude Code headless mode burns ~3x more of the 5-hour subscription limit per token, tests show
Bitter-Truck1049 · reddit · 2026-09-29
A Reddit user ran the same mid-sized Python monorepo refactor six times across four ways of running Claude Code (same commit, model Opus 5, effort medium) and compared /usage deltas.
Key findings
- Interactive use (VS Code extension, terminal CLI): 1.6–1.8% of the 5-hour limit per $1 of API-equivalent tokens
- Headless (Python Agent SDK, claude -p): 5.2% per $1 — roughly 3x the limit per token
- Headless runs actually used fewer tokens and smaller context, yet consumed 2x the quota per run; interactive CLI matched VS Code, so the split is interactive vs. headless, not the interface
- Custom system prompts and SDK settings ruled out: claude -p with defaults showed the same 3x ratio
Practical takeaway: a headless task costing $18 at API prices eats 90% of a 5-hour window; the same work interactively takes 30%. Subscription users running headless agents should budget accordingly.
Caveats: six runs, one task/model/account (Pro), ±1% /usage readings, and the actual accounting signal is unknown.
More from coding & agent
- Google Research open-sources PaperOrchestra, a multi-agent framework for writing AI papers — google-research · 2026-11-16
- ffmpeg-skill gives coding agents local video editing via 42 FFmpeg tools and MCP — kajisho5 · 2026-11-16
- Agents Now Reply to Slack Messages on Their Own, Resolving a Prod Incident Ghost-to-Ghost — iamrobotbear · 2026-11-16
- Engineer's Interview Prep Method: Use Claude to Find Which Answers Undersell Your AI Work — brandon_galang · 2026-11-16
- DeepSeek open-sources its app frontend "chat-demo", already at 3.9k stars — NielsRogge · 2026-11-16
- lemire: AI's real impact on coding is expanding the engineer's 'cognitive radius', not speed — lemire · 2026-11-16