[Bug] Unexpected rapid token exhaustion with Sonnet 4.6 (medium effort) after recent update

Title

[Bug] Unexpected rapid token exhaustion with Sonnet 4.6 (medium effort) after recent update

---

Bug Description

I am a Max plan user and have been consistently using Sonnet 4.6 with medium effort.

After the recent update (following the token-related fix), I started experiencing significantly faster token consumption. The same usage patterns that previously lasted much longer now exhaust tokens noticeably quicker.

This is the first time I am observing this behavior, and it appears inconsistent with prior model behavior.

---

Environment Info

• Platform: darwin (macOS)
• Terminal: zed
• Claude Code Version: 2.1.121
• Plan: Max
• Effort: medium
• Feedback ID: 403fc440-b55e-4e95-a329-cc7b1ab1457e

---

Expected Behavior

• Token consumption should be relatively consistent for similar usage patterns
• medium effort should not result in sudden or disproportionate increases in token usage
• Recent updates should not significantly change consumption without clear documentation

---

Actual Behavior

• Tokens are depleted significantly faster than before
• Similar prompts and workflows now consume noticeably more tokens
• The increase is large enough to impact usability within a single session

---

Reproduction Notes

While this does not occur with a single isolated prompt, it consistently happens during normal usage patterns (interactive coding / iterative prompts).

Steps:

1. Use Sonnet 4.6 with medium effort
2. Run iterative prompts in a coding workflow
3. Observe token depletion rate compared to previous sessions

---

Additional Context

• This behavior started after the most recent update related to token handling
• No major change in prompt size or workflow complexity
• Could be related to:

• effort scaling changes
• internal reasoning/token accounting
• or a regression in token usage behavior

---

Errors

[{"error":"Error: NON-FATAL: Lock acquisition failed for /Users/ferhatkatar/.local/share/claude/versions/2.1.121 (expected in multi-process scenarios)\n    at p66 (/$bunfs/root/src/entrypoints/cli.js:2769:2177)\n    at YY8 (/$bunfs/root/src/entrypoints/cli.js:2769:1257)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-28T17:23:28.537Z"}]

(Note: This error appears to be non-fatal and may not be directly related, but included for completeness.)

---

Impact

• Reduced usability of the Max plan
• Increased cost / faster quota exhaustion
• Harder to predict usage for longer workflows

---

Request

Could you clarify:

1. Is this expected behavior due to recent model or tokenizer changes?
2. Or is this a regression in token usage or effort scaling?

If expected, clearer guidance on token usage changes with medium effort would be very helpful.