[GPT-5.6 Luna xhigh] Severe weekly quota drain, long stalls and capacity errors since Sep 14

Summary

Since September 14, 2026 (Europe/Warsaw), Codex using GPT-5.6 Luna with Very High / xhigh reasoning has behaved very differently from my normal baseline.

In roughly one day, I have consumed almost my entire weekly Codex allowance. Under my normal workflow, the same weekly allowance usually lasts me about a full week.

At the same time, Luna xhigh has become much less reliable: it can spend a long time reasoning without producing useful progress, tasks stall or disconnect, and I sometimes receive Selected model is at capacity. Please try a different model.

Subscription

• ChatGPT Plus
• Codex access through my ChatGPT subscription

Model / reasoning

• Model: GPT-5.6 Luna
• Reasoning effort: Very High / xhigh

What issue are you seeing?

Starting on Sep 14, 2026, I observed all of the following while using Luna xhigh:

• Luna can remain in a reasoning/thinking state for a very long time without making proportional progress.
• Tasks sometimes stall, stop responding, or disconnect.
• Requests can fail with:

Selected model is at capacity. Please try a different model.

• The weekly usage allowance is being depleted dramatically faster than before.
• In approximately one day, almost the entire weekly allowance was consumed.
• Historically, with my normal development workflow, the weekly allowance lasted around seven days.

The sudden change in both reliability and quota consumption started at approximately the same time.

Why this looks abnormal

My workflow has not intentionally changed from a week-long usage pattern to one that should consume nearly the entire allowance in one day. The scale of the difference is large enough that normal run-to-run variation does not seem like a sufficient explanation.

Because Luna xhigh is also stalling and producing capacity errors, I am concerned that usage may be amplified by one or more of the following:

• repeated retries after backend/capacity failures;
• repeated reasoning/model turns while a task appears stalled;
• context replay or compaction loops;
• incorrect model weighting;
• incorrect quota attribution/accounting;
• account-specific routing/provisioning problems.

I cannot determine the root cause from the client side.

Steps to reproduce / observed sequence

1. Sign in to Codex with a ChatGPT Plus account.
2. Select GPT-5.6 Luna.
3. Set reasoning to Very High / xhigh.
4. Use Codex for normal development work.
5. Observe some requests spending a long time reasoning or stalling.
6. In some cases observe Selected model is at capacity. Please try a different model.
7. Check weekly usage before and after the work period.
8. Observe the weekly allowance falling dramatically faster than the historical baseline.

Expected behavior

• Luna xhigh should complete or fail requests cleanly rather than remain stuck for long periods.
• Capacity failures should not trigger hidden/repeated work that consumes disproportionate quota.
• Weekly usage should remain broadly consistent with comparable historical usage.
• A workload that normally fits within approximately one week should not suddenly consume almost the entire weekly allowance in about one day without a clear explanation.

Actual behavior

• Long reasoning periods with little/no visible progress.
• Stalled/disconnected tasks.
• Intermittent capacity errors.
• Nearly the entire weekly allowance consumed since Sep 14 despite a historical baseline of roughly one week for comparable usage.

Requested investigation

Please correlate my account's server-side telemetry for the affected Sep 14-15 window and check:

1. quota charged specifically to GPT-5.6 Luna xhigh;
2. retries caused by capacity/backend errors;
3. hidden/background model calls;
4. repeated context replay or compaction;
5. model weighting and cached-input treatment;
6. whether any requests were routed to the wrong quota bucket or entitlement state;
7. whether stalled/failed requests were still consuming quota repeatedly.

If the server-side accounting shows that quota was consumed incorrectly because of retries, routing, background work, or another bug, please restore the incorrectly consumed allowance.

It would also be very useful to expose per-task/per-model quota attribution so users can see exactly what consumed the weekly allowance.

Related reports

• #39818 — GPT-5.6 Luna weekly limit dropping abnormally fast
• #41220 — abnormal quota depletion / accounting cross-report tracker
• #43337 — capacity errors despite available allowance

This report may overlap with those issues, but I am filing it because the combination of Luna xhigh stalls + capacity errors + near-total weekly quota depletion within about one day appeared together beginning Sep 14, 2026.

No files, tests, or code entry points are named. Start by reviewing related issues #39818, #41220, and #43337, then correlate server-side telemetry for the Sep 14–15 window; done means identifying whether quota was incorrectly consumed and whether affected allowance can be restored.