Codex Weekly Quota Dropped to Zero? The Bug, the Ranges, and What to Check

You stop using Codex for a few hours, come back, and the weekly quota that said 21% remaining now says 0% — with the reset date moved to a different day. That's not a hypothetical; it's the report in GitHub issue openai/codex#44233, filed September 9, 2026, and labeled by maintainers as both bug and rate-limits. But plenty of "disappearing quota" reports aren't bugs at all — they're the rolling-window mechanics doing exactly what they were designed to do. Here's how to tell which one you're looking at.

The actual bug report

Issue #44233 is unusually well-evidenced for a quota complaint. A ChatGPT Plus user on Windows 11 logged local telemetry showing:

• 14:52 CEST: 5-hour window at 100% used; weekly limit at 79% used (~21% remaining); server-reported weekly reset: September 15, 2026, 08:17 CEST.
• Between 14:52 and 19:08: no agent work, no token generation — the machine was idle.
• ~19:08: the 5-hour allowance showed fully restored before its reported reset time, while the weekly allowance showed 0%.
• Next request: rejected immediately with usage_limit_exceeded, and the retry message named September 12 as the weekly reset — three days earlier than the server had previously reported.

Two anomalies, then: ~21 percentage points of weekly allowance vanished with no attributable usage, and the weekly reset date silently moved backward. The reporter links it to a broader tracker of quota/accounting reports (#41220) and similar recent symptoms (#44205, #44206). As of this writing the issue is open. We can't tell you what the server-side ledger says — that's precisely what the issue asks OpenAI to audit.

What OpenAI actually publishes (ranges, not numbers)

This is where most confusion starts: OpenAI does not publish fixed Codex quotas. It publishes estimated ranges of local messages per rolling five-hour window, and says out loud that they're estimates. Per OpenAI's pricing page as read by third-party trackers in September 2026:

Two footnotes matter more than the table. First, the meter counts tokens, not messages — one giant multi-file task can cost more than fifty quick edits, which is why two people on the same plan report wildly different experiences. Second, a weekly cap sits on top of the 5-hour window; OpenAI's page says "weekly limits may also apply" without publishing a weekly number or a fixed reset day. Neither fact is an excuse for quota evaporating while idle — but together they explain most non-bug reports of it.

Bug or mechanics? A quick triage

Ask three questions:

1. Did the number change while you were actively working? Quota declining during use is the meter running — expected. Quota dropping while idle (as in #44233) is either a real accounting bug, usage from another surface on the same account (cloud chats, ChatGPT Work usage drawing from the same allowance), or delayed reconciliation. All three are worth reporting; only the first is OpenAI's to fix.
2. Did the reset date move? A 5-hour reset time shown in the dashboard is authoritative for that window. A weekly reset date that shifts without a documented entitlement change is exactly what #44233 documents — attach telemetry and file it.
3. Were you actually near the weekly cap? Several consecutive days of heavy agentic runs are the normal way people meet the weekly limit. The 5-hour window refills and tricks you into thinking you're fine; the weekly cap is the one that only appears after the third long day.

What to check, and what to do

Check /status first. Inside a Codex CLI session, /status shows current token usage and reset times without leaving the terminal; /usage shows daily/weekly/cumulative activity, and the ChatGPT usage dashboard shows the same meters with exact reset times. Don't diagnose from a remembered number — the ranges move.

If it looks like mechanics: wait out the 5-hour window, downshift to a cheaper model (Terra or Luna instead of Sol) so what remains lasts longer, tighten prompts so tasks close in fewer turns, or buy additional credits — Plus and Pro can top up without changing plan. We covered the larger strategy in Codex usage limits explained.

If it looks like #44233: capture telemetry before it rolls over — screenshot /status, note exact timestamps and timezones, and open an issue on openai/codex referencing the tracker. Support won't reset a limit for you. And if you've ever caught yourself maintaining a quota tracker just to know when you'll be blocked next, that's a symptom, not a fix.

See the pricing yourself: meshcode is free to start — download it

The honest bottom line

As of September 2026 there are two truthful sentences about Codex quota, and you need both: OpenAI publishes ranges, not guarantees, on rolling windows with an unpublished weekly cap — and there is a real, labeled, still-open bug where allowance vanished while idle. Knowing which one you're facing turns a bad afternoon into either a two-minute /status check or a well-written bug report. And if you'd rather run your Codex CLI somewhere it shares a workspace with other agents instead of being your only pane, run Claude and Codex side by side — in meshcode, your own CLI keeps its own quotas and its own billing, with no extra meshcode token charge.

👉 Download meshcode — Mac, Windows