[Pro 20x] Effective GPT-6 Astra allowance feels below one third of previous capacity; Sep 14–15 quota evidence

Pro 20x: reported severe reduction in effective GPT-6 Astra allowance and perceived quality regression

Account and environment

• Subscription: ChatGPT Pro 20x, as reported by the user. The live Codex account endpoint confirms planType: pro; it does not expose the 20x multiplier in the returned snapshot.
• Installed CLI: codex-cli 0.154.0.
• Local diagnostic environment: Fedora 44, Linux 7.2.5-200.fc44.x86_64, x86_64. Desktop app version has not been verified.
• Model recorded locally: gpt-6-astra.
• Timezone for the table below: Asia/Shanghai (UTC+8).

User experience

Around last weekend (approximately September 12–13, 2026; onset is uncertain), my practical usable allowance began to feel like less than one third of what it used to be. Previously I could sustain GPT-6 at Max reasoning; now even High does not provide enough usable capacity for my work. I also perceive a substantial deterioration in model quality.

“Less than one third” is my estimate of practical capacity, not a measured entitlement ratio. No controlled before/after quality example has been supplied yet. Please investigate these reports without assuming the cause in advance.

Verified quota observations

These are account-level quota snapshots recorded in local Codex history, plus a fresh account/rateLimits/read response. All rows refer to the codex bucket and a 10,080-minute (weekly) window.

The next weekly reset returned by the live API is September 19, 2026, 16:44:45 UTC+8. This snapshot returns no secondary window for the main codex bucket, so it does not establish a 5-hour percentage for Astra. The separate Spark bucket must not be used as Astra's quota.

The observed main weekly counter increased by 30 percentage points during the Sep 14 interval and by 14 points during the Sep 15 interval. These are account-level changes across concurrent activity, not consumption attributed to one task.

Local usage evidence and limits

A local review covers 61 rollout files from threads updated since Sep 7. Earlier records predominantly show Astra max; Sep 14–15 records contain both high and xhigh. The recent consumption cannot be attributed solely to High. Local records show input, cached input, output and reasoning token counters; a filtered diagnostic attachment contains those counts, timestamps, session identifiers and quota snapshots.

The counts are local observed token usage, not server billing units. Cached input is a subset of input; reasoning output is a subset of output. This device's records may not cover other devices or all account activity. No local usage records were found for Sep 12–13 in this review, so the exact onset is not established. The observations alone do not prove a threefold change in quota weighting or a backend model downgrade.

Requested investigation

Please perform an account-level quota accounting investigation and determine whether the effective quota weighting or quota accounting for ChatGPT Pro 20x has changed.

1. Verify the account's correct entitlement and explain any changes in effective allowance or model/reasoning weighting around Sep 11–15.
2. Reconcile the recorded quota changes with server-side per-request usage, including input/cache treatment, reasoning output, subagents, retries, failed requests and recovery work.
3. Check whether any quota resets, entitlement changes or account transitions affected the comparison period.
4. Investigate the separately reported quality regression, including the actual model snapshot/routing and effective reasoning configuration. Please state what additional reproducible examples are needed.
5. Provide the findings and an understandable usage breakdown; correct the allowance if erroneous accounting is identified.

Related public report: https://github.com/openai/codex/issues/43967. This is a separate account's complaint with its own local measurements; a common root cause has not been established.

Codex feedback submission

Feedback tracking ID: no-active-thread-01a0a416-21c5-7483-994b-60d6733ab2d0

Submitted successfully through Codex app-server feedback/upload with a filtered local usage diagnostic JSON attachment. Full conversation logs were not uploaded. This is an account-level report, so the feedback tracking ID has a no-active-thread- prefix. The diagnostic attachment includes session identifiers for correlation.

Start with the live account/rateLimits/read response and the filtered diagnostic JSON attached through feedback/upload. Compare the account-level quota snapshots with the 61 local rollout records, while keeping the separate Spark bucket and the related issue distinct. Done means a server-side usage reconciliation, entitlement and routing findings, and a clear explanation or correction if accounting is wrong.