Usage cap: single ~30k-token prompt (no subagents) burned 38% of 5h Max 20x limit on Opus 4.7

Summary

A single Claude Code prompt on Opus 4.7 consumed ~38% of my 5-hour Max 20x usage cap in under 20 minutes. The prompt produced a written plan of roughly 30,000 tokens, with no subagents dispatched and no parallel tool fan-out.

Environment

• Plan: Max 20x ($200/mo)
• Model: Opus 4.7 (claude-opus-4-7, 1M context)
• Client: Claude Code CLI
• Date: 2026-05-25
• Duration: under 20 minutes
• Cap burn: ~38% of 5-hour rolling window

What I expected

30k tokens of output (~a medium-length design doc) consuming a roughly proportional slice of the cap; on the order of single-digit percent, not ~38%.

What happened

~38% of the Max 20x cap gone on one linear prompt. No subagent dispatches, no Agent tool calls, no parallel work.

What I can't see

There's no user-facing breakdown of where the budget went. Possible hidden consumers:

• Extended thinking tokens at a different multiplier
• Cached read tokens being re-billed
• Tool-result echoes counted against the cap
• Something else entirely

Happy to share my Claude Code session ID privately via support if it helps reproduce.