Token usage regression: 3-4x increase over past 2 weeks

Description

Experiencing a significant token usage regression over the past 2 weeks. Simple commands that previously consumed minimal tokens are now burning through budget at 3-4x the previous rate.

Impact

Before (2 weeks ago):

• Could run 3-4 parallel Claude Code sessions
• Full day of productive work on one battery cycle

Now:

• Single session for ~1 hour
• 15% battery consumed by two trivial commands (rtk gain + ls)

Specific Example

Simple ls -la command:

• Actual tokens used: ~6,000 tokens (3% of 200K budget)
• Expected tokens: ~1,000 tokens (500 tool result + 300 response + 200 overhead)
• Excess overhead: ~5,000 tokens that should be cached

Evidence

User has RTK (token tracking tool) showing historical baseline vs current usage. The 3-4x regression is consistent across all command types.

Context

User mentioned "anthropic have admitted to a bug causing extra token usage" - this report may be related to that known issue. If there's an existing ticket, please link.

Environment

• Model: claude-sonnet-4-5-20250929
• Platform: macOS (Darwin 25.3.0)
• Regression started: ~2 weeks ago (mid-March 2026)

Request

This is blocking productive work. Is this the known bug? What's the ETA on a fix?