[Bug] Rapid token usage spikes with Opus model in Claude Code sessions

Bug Description
Bug Report for Claude Code Usage Calculation
Issue: Rapid token usage spikes (15-16% jumps within minutes) starting April 4, 2026 at ~13:28 IST

Details:
Max Plan individual subscription
Using Claude Code with Opus model, "high" effort setting
Session 1 at 2026-04-04 (from 1329 to 1518 IST): Ran out of tokens in ~1hr 50min (normally lasts much longer)
Session 2 at (2026-04-04 (from 2107 IST ongoing): Hit 74% usage in just 15 minutes of session start (with 15-16% spikes in 1-2 minutes. Jumped from 26% to 42% then to 58% then to 74%). Closed all sessions and stopped using.

Had changed settings.json with this key yesterday:
"CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
Have now removed this key.

No changes to usage patterns - same long conversation without auto-compact that worked fine previously.

Suspected Technical Issue:
Potential cache_read_input_tokens miscalculation where internal API calls from server-side tools are incorrectly counted against user limits.
Investigation Request:
Please examine if cache_read_input_tokens from multiple internal Claude Code operations are being incorrectly added to usage calculations, causing inflated token consumption reporting, or if there is some other issue.

Environment Info

• Platform: win32
• Terminal: windows-terminal
• Version: 2.1.92
• Feedback ID: ce6bd468-ad45-47a3-82db-3c3021533b32

Errors

[]