[BUG] Max 20x quota drains to 100% in ~15 min while local usage is minimal (~$0.25 API-equivalent)

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?

My Max 20x plan usage limit is being consumed at a rate wildly disproportionate
to my actual usage. The 5-hour window can reach 100% in ~15 minutes of light,
single-session work.

Measured evidence (2026-07-23):

• I opened a fresh session (new 5h window, /usage showed "Resets in 4 hr 59 min",

so the window had just started).

• After a single short exchange, /usage already showed 10% used.
• Local consumption measured from the session transcripts (via ccusage) for that

same period: ~41k total tokens — 20k cache read, ~20k cache write, 629 output
tokens (~$0.25 API-equivalent). Two exchanges, no heavy tooling.

• A few days earlier (2026-07-18), the same pattern saturated the window to 100%

in ~14 minutes.

Things I ruled out:

• Only one claude process running on the machine (verified with ps aux).
• No scheduled cron jobs / cloud routines under my account.
• Only one transcript file was modified in the last 90 minutes across all

local projects (~/.claude/projects).

• No other device or claude.ai web/mobile session active at the time.

Model: claude-opus-4-8 with 1M context ([1m]).
Question: is the 1M-context model weighted at a premium rate against plan
limits? If so, at what multiplier? This is not documented anywhere I can find,
and even a premium multiplier would not obviously explain 10% of a Max 20x
window for a ~41k-token exchange.

What Should Happen?

Plan usage consumption should be proportional to actual token usage. A ~41k-token
exchange (mostly cache reads, 629 output tokens) should consume a negligible
fraction of a Max 20x 5-hour window — not 10%.

If the 1M-context model consumes quota at a premium rate, that rate should be
documented, and ideally surfaced in /usage or /status so users can understand
and predict their consumption. Users should also have a way to see a breakdown of
what consumed their quota (per session / per device / per surface), so that
unexpected drains like this can be diagnosed.

Error Messages/Logs

Steps to Reproduce

1. Be on a Max 20x plan, using Claude Code CLI with model Opus 4.8 with 1M

context (claude-opus-4-8[1m]).

1. Wait for the previous 5-hour usage window to reset (or start fresh).
2. Open a single new Claude Code session in a terminal.
3. Send one or two short prompts (in my case: a couple of conversational

messages plus a few lightweight read-only tool calls — ~41k total tokens,
629 output tokens, as measured from the local session transcript).

1. Run /usage.

Result: the freshly opened 5h window ("Resets in 4 hr 59 min") already shows
10% used. Continuing to work normally saturates the window to 100% in ~15
minutes (observed twice: 2026-07-18 and 2026-07-23).

Expected: a ~41k-token exchange should consume a negligible fraction of a
Max 20x 5-hour window.

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.218

Claude Code Version

2.1.218

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_