[BUG] Pro 5x weekly allowance drains unexpectedly, including idle periods and after visible tasks finish

Summary

I’m seeing a large regression in effective Codex weekly allowance on ChatGPT Pro 5x ($100). The weekly meter now drops much faster than my previous GPT-5.6 Sol workflow, and I have also observed allowance decreasing during periods when I had no intentionally running Codex or Work task.

This appears related to the broader symptom family tracked in #41220, but I’m filing a concrete account/workflow report because I have local rollout/token evidence and server-provided rate-limit snapshots available for maintainer/support correlation.

Environment

• Plan: ChatGPT Pro 5x ($100)
• Codex: 0.153.4
• Platform: macOS x86_64
• Primary surfaces: Codex Desktop / previously VS Code Codex
• Work: no active Work tasks during the idle-drain observations
• Models used recently: GPT-5.6 Sol and GPT-6 Astra
• Astra was not being used continuously during all observed drain

Previous baseline

Before the recent Astra rollout / current metering behavior, I could regularly run roughly four GPT-5.6 Sol Extra High tasks in parallel and a heavy working day would typically consume about 30–40% of the weekly allowance.

The current behavior is materially worse even when I use Sol and when Astra usage is limited.

Current symptoms

1. Weekly allowance drops much faster than the prior baseline for comparable work.
2. I have observed the weekly percentage decrease while I believed no foreground Codex task or Work task was running.
3. I previously observed spawned Astra subagents continuing work after the parent task had already produced its final report. The visible task appeared complete, but background activity continued.
4. The effective usable weekly allowance now feels substantially smaller than before even when I switch back to GPT-5.6 Sol.
5. I cannot reconcile the user-visible meter with the amount of useful foreground work being performed.

Local evidence collected

I collected a sanitized evidence bundle containing:

• Codex version and codex doctor output;
• token counters from local rollout JSONL grouped by model / effort / source;
• daily token aggregation;
• server-provided rate-limit snapshots already written into Codex rollouts;
• local process snapshot;
• no prompt bodies, credentials, API keys, or private project content.

The same evidence bundle has also been sent privately to OpenAI Support so the account-specific server-side meter can be audited without publishing identifying information here.

What I would like maintainers to verify

Please correlate the affected account’s server-side metering for the recent window against:

• foreground model requests;
• subagent/background activity;
• retries / polling / heartbeat re-entry;
• compaction or repeated context replay;
• model + reasoning effort;
• the quota bucket actually charged;
• any entitlement or weighting changes affecting Pro 5x;
• any usage generated after the parent task had already returned its final answer.

If the meter is charging unintended background work or is otherwise mis-accounting usage, please reconcile/restore the incorrectly consumed allowance.

Expected behavior

• Weekly allowance should correlate with actual foreground/background work that is attributable and visible to the user.
• A task that has produced its final result should not leave hidden agents or background loops consuming quota indefinitely.
• Idle periods with no active task should not consume weekly allowance.
• Comparable Sol workflows should not suddenly consume a dramatically larger fraction of the same weekly bucket without a documented reason.

Related

• #41220 — meta tracker for abnormal quota depletion / accounting inconsistencies

I can provide timestamps and additional sanitized local evidence if maintainers need a narrower reproduction window.

No repository file or test is named. Start with the sanitized evidence bundle, codex doctor output, rollout JSONL token counters, rate-limit snapshots, and process snapshot; compare foreground requests with background activity, retries, polling, compaction, and post-completion work. Done means establishing whether usage is being mis-accounted and documenting a reproducible window for maintainer or support correlation.