[Usage] Pro 20x weekly allowance can be exhausted in one day — investigate capacity and agent overhead

Summary

I am a long-term ChatGPT and Codex user on the Pro 20x plan. In my recent experience, one day of legitimate, complex, long-horizon Codex agent work can consume the entire weekly allowance.

Please do not treat this only as "a power user using the product too quickly." The user-visible problem is that the highest individual subscription tier may no longer sustain a week of serious agentic work. This warrants a joint investigation of both effective plan capacity and system-side usage efficiency.

Environment

• Subscription: ChatGPT Pro (20x)
• Client: Codex Desktop / ChatGPT app 26.818.31338 (build 6892)
• Platform: macOS 26.5.1, Apple silicon
• Workload: legitimate software-engineering tasks involving long contexts, tools, and agent orchestration

What should be investigated

At least three explanations need to be separated rather than assumed:

1. Effective capacity changed — the usable Pro 20x allowance may have decreased or may be provisioned incorrectly for some accounts.
2. Per-task consumption regressed — the current model/agent stack may consume materially more counted usage to complete the same class of work.
3. System overhead is consuming the allowance — context compaction, repeated context reconstruction, tool output replay, retries, polling/wait turns, or orchestration overhead may be counted inefficiently or incorrectly.

A weekly aggregate percentage is not sufficient to distinguish these causes.

Why this matters

The concern is not a request for another promotional reset. Temporary reset credits may relieve the symptom, but they do not explain why a Pro 20x allowance can disappear in one working day.

If the highest individual tier cannot sustain more than a day of serious use, the practical usability of lower tiers also deserves review. Users need to know whether they are paying for productive task execution or for avoidable system overhead.

Requested actions

Please:

1. Compare current Pro 20x effective capacity with prior releases and disclose whether the entitlement or model weighting changed.
2. Measure usage per completed task, not only aggregate tokens or credits.
3. Break down usage into uncached input, cached input, output, reasoning, compaction/reconstruction, tool traffic, retries, waits/polling, and subagent orchestration.
4. Investigate regressions across long sessions and tool-heavy or multi-agent workflows.
5. Add per-task usage visibility and warnings for abnormal consumption, repeated compaction, retry loops, or low-progress background activity.
6. Correct and restore allowance where provisioning or accounting errors are confirmed.

Evidence boundary

This report describes the user's observed experience. It does not claim that a quota reduction or billing defect has already been proven. Only OpenAI can verify the backend entitlement, weighting, and accounting data.

No account identifiers, private prompts, repository names, local paths, credentials, or proprietary project content are included here. Sanitized diagnostics can be provided privately if maintainers request them.

Disclosure from the Codex assistant

This feedback was raised by a Pro 20x user and organized in English by their Codex assistant. Based on the reported pattern and the related public reports below, the assistant assesses that the issue warrants formal investigation. The assistant is not an OpenAI spokesperson and has no access to OpenAI's backend billing ledger.

Related public reports

• #38157 — Pro 20x accounts appear to receive Pro 5x effective capacity
• #38728 — weekly quota meter accelerated approximately 2.4x during a bounded interval
• #35259 — repeated model-mediated wait/status polling consumed substantial credits
• #37090 — abnormal consumption and repeated context-compaction loops

No repository file, test, or entry point is mentioned in the report, so there is no payload-grounded starting location to read or run. Done would require maintainer-verified measurements separating entitlement, per-task consumption, and system overhead, followed by correction if an accounting or provisioning defect is confirmed.