[Codex app] GPT-6 Astra Ultra consumed ~50% of a Pro weekly quota in ~80 minutes across two Goal runs

What version of the Codex App are you using (from the “About” dialog)?

26.901.51231 — released September 5, 2026.

The About dialog identifies the application as ChatGPT — Powered by Codex & OWL; the affected workflow is in the app's Codex interface.

What subscription do you have?

ChatGPT Pro — $200/month, 20x Codex usage tier.

What platform is your computer?

• macOS Tahoe 26.6.2
• MacBook Pro, 16-inch, Apple M5 Max
• 128 GB memory

Model and workflow

• GPT-6 Astra Ultra
• Codex desktop app
• Goal workflow
• Goal 1 displayed: Worked for 1h 1m 39s
• Goal 2 displayed: Worked for 18m 24s
• Total displayed Goal work: approximately 1h 20m

What issue are you seeing?

Based on the weekly-allowance indicator before and after these runs, approximately 50% of my entire weekly quota was consumed while completing these two Goals.

This consumption rate is extraordinarily high for a $200/month Pro subscription. At the observed rate, similar continuous use would exhaust the full weekly allowance in roughly 2 hours and 40 minutes. In practical terms, using the newest model continuously for only two to three hours could consume an entire week's quota.

I understand that Astra Ultra and agentic workflows may carry a higher usage weight, but losing half of a 20x Pro weekly allowance after about 80 minutes of displayed Goal work is not reasonably sustainable. It may indicate excessive model weighting, unexpected agent/context overhead, incorrect entitlement application, or a usage-accounting problem.

I do not have access to exact per-run token counts or the server-side accounting breakdown, so please verify the attribution using the submitted diagnostics rather than treating the 50% figure as token telemetry.

What steps can reproduce the bug?

1. Use the Codex interface in the macOS app with a ChatGPT Pro 20x subscription.
2. Select GPT-6 Astra Ultra and note the remaining weekly allowance.
3. Run a Goal task for approximately one hour.
4. Run another Goal task for approximately 18 minutes.
5. Compare the weekly allowance afterward.
6. In this case, the two Goals totaled about 1h 20m, while approximately 50% of the weekly quota was consumed.

What is the expected behavior?

Weekly usage on the highest paid consumer tier should be predictable and sufficient for sustained development work. A pair of ordinary Goal runs totaling around 80 minutes should not consume approximately half of the weekly allowance without a clear, user-visible explanation.

Please investigate:

• Whether the account's 20x Pro entitlement was applied correctly.
• How much usage was attributed to each Goal and any associated subagents.
• Whether retries, duplicated metering, context compaction, caching, tool calls, or hidden recovery work were counted unexpectedly.
• Whether the newer app/model combination introduced a regression in usage efficiency or accounting.
• If this consumption is intentional, whether the quota, per-model weighting, and per-Goal usage can be made transparent in the UI.

Additional information

Feedback ID: 01a07f17-ba47-79b3-999c-1263f888d09f

The in-app feedback submission completed successfully with diagnostic logs enabled.

Related earlier report from the same account: #43029. That report observed roughly 30% weekly-quota consumption during a roughly one-hour Astra Ultra task on app version 26.901.31953. This is a new measurement on version 26.901.51231, with a new diagnostic Feedback ID, and the observed depletion is even more severe.

Screenshots document the app version, selected model, both Goal durations, and feedback confirmation. The device serial number and unrelated project details are intentionally omitted.

No repository file or test is named. Start by reviewing diagnostic feedback ID 01a07f17-ba47-79b3-999c-1263f888d09f alongside related issue #43029, then compare the reported Goal durations with entitlement and usage attribution. Done means the discrepancy is reproduced or explained with per-run accounting, and any accounting or UI change has appropriate regression coverage.