GPT-5.6 Sol suddenly exhausts Codex 5-hour and weekly allowances far faster than before

What subscription do you have?

I have reproduced this across:

• ChatGPT Pro 20x
• ChatGPT Plus

Which model were you using?

GPT-5.6 Sol.

This report is specifically about Sol. I previously used GPT-6 Astra, but switched away from it because of higher usage. The abnormal behavior described below is occurring on GPT-5.6 Sol.

What issue are you seeing?

Codex usage consumption has suddenly increased by a very large amount for the same kinds of development tasks I have been running for roughly the past month.

The change is recent and consistent.

On my Pro 20x account:

• GPT-5.6 Sol can now consume essentially my entire weekly Codex allowance in roughly 12 hours.
• Previously, comparable sustained workloads lasted dramatically longer.

On a separate Plus allowance:

• I started a GPT-5.6 Sol task.
• After approximately 14 minutes, the entire 5-hour allowance was exhausted.
• This was effectively one task/message, not many hours of interactive use.

I have also observed the usage indicator jump directly from approximately 6% remaining to 0%, rather than visibly stepping down through intermediate values.

This is not an occasional outlier. It is now happening consistently when I send tasks/messages to Codex.

The important comparison is longitudinal: the tasks and workflow are materially similar to what I had been running successfully for about a month, but the amount of subscription usage consumed has changed dramatically.

I understand that the "5-hour" limit is a usage window rather than five literal hours of runtime. The issue is the sudden change in effective throughput for the same model and comparable work.

What steps can reproduce the bug?

1. Use GPT-5.6 Sol on Codex.
2. Run the same type of sustained development task that previously consumed usage at a normal/predictable rate.
3. Observe that current usage falls far faster than it did previously.
4. On Plus, a single Sol task can exhaust the full 5-hour allowance in roughly 14 minutes.
5. On Pro 20x, sustained Sol usage can now exhaust the weekly allowance in roughly 12 hours.

Observed on 2026-09-20, Europe/London timezone.

What is the expected behavior?

For materially similar GPT-5.6 Sol workloads, usage consumption should remain reasonably consistent over time unless OpenAI has intentionally changed the allowance or model weighting.

A single ordinary development task should not unexpectedly exhaust an entire 5-hour Plus allowance within ~14 minutes without a clear explanation of the model/tool activity responsible.

What I would like investigated

Please inspect affected Codex usage accounting and recent regressions involving:

• unexpectedly high numbers of model requests generated from one user message;
• repeated replay of large cached contexts;
• cached-input quota weighting/accounting;
• context compaction loops;
• background/system-generated Codex activity;
• sub-agent or worker activity not obvious in the main conversation;
• delayed quota reconciliation that can cause sudden percentage jumps;
• recent changes to GPT-5.6 Sol subscription quota weighting;
• any regression that would explain a sudden large drop in effective quota throughput.

Please compare server-side model-request/token activity with the amount deducted from the 5-hour and weekly subscription limits.

Related reports

This appears closely related to:

• #45867 — large recent regression/change in effective quota throughput and delayed quota jumps
• #46816 — entire 5-hour allowance consumed in ~20 minutes with many internal model requests and cached-context replay
• #46689 — Sol Medium consuming ~23% of a 5-hour allowance during one relatively small task

I can provide timestamps, screenshots, affected task/session IDs, and local logs privately if maintainers specify the safest way to share them.

No repository file, test, or entry point is identified. Start by comparing server-side model-request and token activity with deductions from the 5-hour and weekly limits, using the timestamps, session IDs, and logs the reporter can provide. Done means isolating whether request generation, cached-context accounting, background activity, quota weighting, or reconciliation explains the abnormal consumption.