GPT-5.6 Sol normal mode appears to consume Codex quota unusually fast on $100 plan

What subscription do you have?

$100/month ChatGPT plan.

Which model were you using?

GPT-5.6 Sol.

Fast Mode is explicitly disabled. This report is about normal mode.

What platform is your computer?

Windows 11 Pro x64.

What issue are you seeing?

Over roughly the last 2–3 days (around August 10–12, 2026), Codex in normal mode has appeared noticeably faster than before, while the included usage/quota is also depleting much faster than expected for comparable ordinary development work.

I have not enabled Fast Mode, so the documented Fast Mode credit multiplier should not explain this behavior.

The abrupt change makes this look like a possible regression in one of the following areas:

• token/context usage;
• cached-input accounting;
• quota/credit calculation;
• automatically injected context, plugins, skills, memory, or tool schemas;
• or server-side entitlement/rate-limit accounting.

There are other current reports describing unusually fast usage depletion, including #38191, so this may not be isolated to one account.

Please investigate whether normal-mode GPT-5.6 Sol usage accounting changed or regressed during the August 10–12 rollout window.

It would also be very useful to expose a per-thread breakdown of input tokens, cached input tokens, output/reasoning tokens, and the corresponding quota/credit charge so users can distinguish genuine model consumption from an accounting problem.

What steps can reproduce the bug?

1. Use Codex on Windows with GPT-5.6 Sol.
2. Confirm Fast Mode is OFF.
3. Perform normal coding tasks similar to tasks used on previous days.
4. Compare the Codex usage/quota meter before and after the tasks.
5. Observe that the available usage appears to drop materially faster than it did before approximately August 10, despite Fast Mode remaining disabled.

What is the expected behavior?

Normal mode should consume usage according to the documented normal-mode token/credit rates. An abrupt increase in depletion should either be documented as an intentional pricing/accounting change or investigated as a regression.

If the backend is correctly charging the usage, the product should provide enough per-thread token and credit detail to explain the difference.

Additional information

I am intentionally not posting account identifiers or private project/session data publicly. Exact session IDs and account-specific usage details can be provided through a private OpenAI support channel if needed.

No repository file, test, or entry point is identified in the report. Start by reproducing the normal-mode usage change on Windows with Fast Mode disabled, then compare quota accounting across the August 10–12 rollout window; done means identifying whether token, cached-input, context, or entitlement accounting changed and documenting the result.