[BUG] Headless `claude -p` (entrypoint `sdk-cli`) uses ~1.8× more of the 5-hour window per token than interactive `cli`, same workload

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

On a Max subscription, the same workload uses about 1.8× more of the 5-hour window when it runs headless (claude -p --input-format stream-json, recorded as entrypoint: "sdk-cli" in the transcripts) than when it runs in the interactive TUI (entrypoint: "cli"). Token usage is the same in both modes.

The Agent SDK help article (https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan) says claude -p "still draw[s] from your subscription's usage limits", and I could not find anything saying it is weighted differently.

After moving the same work from the TUI to claude -p, we hit the 5h limit several times a day with no change in workload, and we paid for extra usage (rejected + isUsingOverage: true) that the same work in the TUI would not have needed.

What Should Happen?

The same tokens on the same model should use the same share of the 5-hour window, whether they come from claude -p or from the interactive TUI. If headless usage is intentionally weighted differently, it should be documented.

Error Messages/Logs

Controlled A/B test, 2026-09-25 10:27–10:41 UTC, Claude Code 2.1.282, claude-opus-5, effort high

Phase        | Test spend (API-equiv.) | Other spend        | five_hour utilization | $ per 1% of 5h
-------------|-------------------------|--------------------|-----------------------|---------------
headless     | $12.74                  | $0.40 (headless)   | 52% -> 60%            | ~1.6 (1.5-1.9)
interactive  | $12.91                  | $1.12 (headless)   | 60% -> 65%            | ~3.0 (2.4-3.9)

Ranges come from the 1% resolution of the utilization reading; they do not overlap.

Historical (same account, all Opus 5, API-equivalent spend per 1% of each 5h window that peaked above 25%):
2026-09-18..22  cli      2.1.267/2.1.270  3.1-4.1 $ per 1%  (8 windows)
2026-09-23..24  sdk-cli  2.1.267          1.6-1.9 $ per 1%  (5 windows)

Steps to Reproduce

1. Create a directory with ~250K tokens of plain text split into files small enough for Read (I used 18 files of ~45K characters of source code).
2. Prepare a fixed prompt list: turn 1 asks Claude to read all files in parallel and reply only "read"; turns 2–41 are short questions about the content answered in one sentence, without tools.
3. Headless phase: run

claude -p --input-format stream-json --output-format stream-json --verbose --permission-prompt-tool stdio --replay-user-messages --append-system-prompt "<one short paragraph>" --dangerously-skip-permissions --model claude-opus-5
and send the prompts one by one as stream-json user messages, waiting for each result. Record rate_limit_event.rate_limit_info.unifiedWindows.five_hour.utilization.

1. Interactive phase: run claude --model claude-opus-5 --dangerously-skip-permissions in tmux, send the same prompts with tmux send-keys, detect turn end with a Stop hook, and read rate_limits.five_hour.used_percentage from the statusline JSON.
2. Before and after each phase, run a tiny probe (claude -p --model claude-haiku-4-5 --output-format stream-json --verbose "reply: ok") to read utilization the same way.
3. Compute each phase's token cost from the session transcripts (message.usage, deduplicated by message id + request id) at list API prices, and subtract any other activity on the account during the phase (measured from the transcripts of every machine on the account).
4. Compare the $ per 1% of the 5-hour window: headless comes out ~1.8× more expensive for the same tokens.

Checks already done:

• Not the auto-mode classifier: both phases used --dangerously-skip-permissions, and 40 of 41 turns made no tool calls.
• Not a larger harness in headless: the first request of a headless session is slightly smaller than an interactive one (median ~66K vs ~88K input tokens on 2.1.267).
• Not missing transcript entries: headless result.total_cost_usd matches the transcript totals.
• Not usage from other machines: measured and subtracted.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.282 (Claude Code) — also observed on 2.1.267

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Non-interactive/CI environment

Additional Information

Related:

• Independent measurement of the same ~1.8× gap from someone else's logs (no controlled test): https://github.com/futin/claude-agents-dashboard/issues/145 (follow-ups #80 and #157 in that repo)
• #81692 (which usage pool claude -p draws from; closed)

I can share the test scripts and the per-turn measurements if that helps.