Pro weekly usage drained 26% for 75.25M mostly cached tokens immediately after Aug 1 global reset

Subscription

ChatGPT Pro

Models used

• gpt-5.6-sol
• gpt-5.5

Issue

Immediately after the global Codex usage reset announced by Tibo on August 1, 2026, my weekly Pro allowance began draining much faster than it did before the reset.

On August 1, a total of 75,250,381 tokens consumed approximately 26% of the weekly Pro allowance, leaving 74% remaining. Comparable usage was behaving normally as recently as July 30-31, before this reset.

The usage was overwhelmingly cached input, so this does not look consistent with the prior effective capacity of the same Pro plan and workflow.

Measured usage for 2026-08-01

Input tokens:        2,724,979
Output tokens:         212,314
Cache write:                 0
Cache read:         72,313,088
Total tokens:       75,250,381
ccusage API-equivalent: $56.15

Approximately 96.1% of the total tokens were cache reads.

Model breakdown:

gpt-5.6-sol
  Input:       1,372,881
  Output:        199,594
  Cache read: 50,787,584
  Total:      52,360,059

gpt-5.5
  Input:       1,352,098
  Output:         12,720
  Cache read: 21,525,504
  Total:      22,890,322

Account/session evidence

Weekly limit remaining: 74%
Next reset: Aug 8 at 14:33
Session ID: 019fb8e8-c151-7e91-b5c1-aebf413a1a3d
Context window: 227K used / 258K

At the observed rate, the full weekly Pro pool would correspond to only about 289M tokens for this same usage mix, which is roughly half the effective capacity observed before the August 1 reset.

Expected behavior

A global reset should restart the same Pro entitlement and accounting behavior. It should not silently reduce the effective weekly capacity by roughly 2x for an unchanged workflow.

Cached input should be charged using the documented cached-input weighting, and internal retries, automatic review/guardian tasks, delayed reconciliation, or duplicate attribution should not unexpectedly consume a large share of the weekly allowance.

Requested investigation

Please inspect the server-side usage ledger for the session above and confirm:

1. The exact credits charged for uncached input, cached input, and output.
2. Whether any internal automatic tasks, retries, subagents, or delayed usage were also charged.
3. Whether the Pro entitlement or model weighting changed after the August 1 global reset.
4. Whether usage was duplicated or reconciled late.
5. Whether incorrectly consumed weekly allowance can be restored.

Screenshots of the token breakdown and the 74%-remaining weekly meter are available if needed.