Codex Usage Evaporated Overnight? OpenAI Acknowledges Bug, Full Reset + Extra Reset

Your Codex quota might not have been used by you—it was consumed by itself.

At 8 AM Beijing time today, the OpenAI Codex team posted an announcement with two key points: full reset of usage limits, and one additional manual reset within 24 hours.

Sounds like a perk? No, it's an admission of fault.

Usage Evaporation: It's Not That You Used Too Much, It's That It Burned Too Much

Over the past few days, Codex users collectively reported an issue: quota consumption was abnormally fast. Even with few tasks run, the usage pool drained quickly.

Some thought their tasks were too heavy, others suspected the model had become more expensive, and a few silently paid for an upgrade—after all, Codex Pro costs $200/month, and once the quota is used up, you have to wait until next month.

But the truth is: your quota was eaten by the system itself.

After investigation, the Codex team confirmed that the rapid consumption was caused by multiple small issues piling up:

• Auto-review too aggressive: Codex automatically reviews code changes, but recently this mechanism became overly aggressive, frequently triggering unnecessary review processes, each consuming tokens.
• A certain change triggered too many subagents: A backend change caused subagent tasks to be over-scheduled; one task might be split into multiple parallel subtasks, each consuming quota independently.
• Background suggestion retries too frequent: When a suggestion generation failed, the system automatically retried, but the retry strategy was overly aggressive, potentially triggering multiple retries for a single failure, each retry counting towards usage.

Each problem alone might not be critical, but together, it's like turning on three faucets simultaneously—your usage pool drops visibly.

Even More Absurd: Auto-review Was Counted as GPT-5.4

If it were just burning a few extra tokens, it would still be a "technical issue." But the Codex team also discovered a bug in the usage report:

The quota consumed by Auto-review was erroneously marked as GPT-5.4 usage.

What does this mean? GPT-5.4 is the most expensive model tier in Codex. If you saw abnormally high GPT-5.4 consumption in your usage report, it might not be because you actively selected GPT-5.4—it was Auto-review quietly "spending" on your behalf.

You thought you were saving, but the system charged you at the highest rate.

What Was Fixed

The Codex team has completed the following fixes:

1. Reverted the change that triggered too many subagents—cutting off overscheduling at the source.
2. Fixed suggestion scheduling behavior—reducing unnecessary suggestion generation.
3. Fixed duplicate generation issue—preventing the same content from being generated multiple times.
4. Fixed retry behavior—no longer retrying excessively after failure.
5. Corrected usage report error—Auto-review is no longer mislabeled as GPT-5.4.

All hotfixes have been deployed to the CLI, desktop app, and backend. This means when you open Codex now, new usage data will be more accurate, and actual consumption will decrease.

What You Need to Do

Three things:

First, check your usage report. If GPT-5.4 usage was abnormally high before, you should now see more accurate data.

Second, wait for the automatic reset. The team promises to fully reset all users' usage limits within the next hour. No action needed—it will be applied automatically.

Third, manual reset. Within 24 hours, you will also receive an additional manual reset allowance. If after the reset you still feel the quota is insufficient, you can manually reset once more.

This Isn't the First Time

Codex's usage issues have not caused user dissatisfaction for the first time. Last month, ChatGPT Pro users woke up to find themselves downgraded to Free users, and even CEO Sam Altman's account was affected. That time it was a subscription tier glitch; this time it's abnormal usage consumption.

The nature of both problems is the same: users lack transparency and control over their account status.

You pay money, but you don't know where your money went. You select a model, but the system might be switching it behind the scenes. Your quota runs out, but you can't tell whether it was your own usage or the system's burning.

This time OpenAI responded relatively quickly—about a day from user feedback to announcement. But "quick response" and "shouldn't have happened" are two different things. Pro users paying $200/month should not need to investigate "why my quota evaporated so fast" on their own.

A Reminder

If you are a heavy Codex user, it's advisable to develop a habit: take regular screenshots of your usage report.

Not because of good memory, but because when the system has a bug, you need evidence to prove "I didn't use this." Today's fix shows that even Codex's usage statistics can be wrong—the numbers you see may not reflect reality.

Fortunately, OpenAI admitted it this time. What about next time?