Excessive token consumption on Opus 4.6 — thinking tokens disproportionate to output

Summary

Simple tasks (3 v8 ignore annotations + ~20 unit tests = 3 commits) consumed 55% → 89% of a 5-hour session quota on Claude Code with Opus 4.6 (1M context).

Evidence from transcript

"Churned for 3m 46s" → produced: nothing (waiting for user input)
"Crunched for 1m 39s" → produced: "Commit efcf14f2. D'autres fichiers à couvrir ?"
"Churned for 37s" → produced: 3 lines of text
Agent Sonnet: 56 tool uses for branch coverage tests → interrupted by user

Quota impact

• Before: 55% session quota used
• After (~20 min of work): 89% session quota used
• Context window: 58% filled, 38h58m accumulated session time
• Paid overage: 44.62€ / 45€ limit (99% consumed)

Screenshots

(See attached — Settings > Usage page showing 72% then 89% shortly after)

!Usage 72%
!Usage 89%

Note: Screenshots will be added in a follow-up comment as GitHub CLI does not support image uploads.

Questions

1. Is it expected that "thinking tokens" (churning/crunching) consume quota even when no output is produced?
2. Does cost scale non-linearly with context size (38h of conversation)?
3. When a Sonnet sub-agent makes 56 tool calls, does each tool call also cost Opus tokens on the parent?

Environment

• Claude Code CLI (latest)
• Model: Opus 4.6 (1M context)
• Plan: Max with paid overage enabled
• OS: macOS Darwin 24.6.0