Skip to content

Never redeem a Codex banked reset before reading this: the global-reset math

Updated October 9, 2026 · first published October 9, 2026

Quick answer: On October 2, 2026 we hit 100% of our Codex weekly limit at 08:48 UTC and redeemed a banked reset a minute later. At 21:10, with 14% of the new window used, OpenAI pushed a global reset to every...

On October 2, 2026 we hit 100% of our Codex weekly limit at 08:48 UTC and redeemed a banked reset a minute later. At 21:10, with 14% of the new window used, OpenAI pushed a global reset to every account. We got a fresh window, but the 86% we had not used of the banked one was simply gone: about $150 of API-priced Codex. Here is how to avoid doing the same.

How banked and global resets work

A global reset zeroes the Codex gauge for every account at once, usually with a post from OpenAI and no warning. A banked reset is a refill OpenAI grants into your account to redeem when you choose. According to OpenAI's help center, redeeming one refreshes your Codex usage windows and moves your weekly reset date. In our case, a later global reset did not credit back the unused part of a redeemed banked reset.

OpenAI resets Codex more often than you think

Public Codex usage resets per month, Sep 2025 to Oct 8, 2026 (codex-resets.com)024681012141Sep 25Oct2Nov4DecJan 26Feb6Mar6Apr3May5Jun12Jul10Aug7Sep3Oct
Public Codex usage resets per month, Sep 2025 to Oct 8, 2026 (codex-resets.com)

The tracker at codex-resets.com logs 59 public resets from September 2025 to October 8, 2026. The pace has changed a lot. Early on there were months with none, and one gap of 67 days. Since July, resets have come every few days: the median gap since August 1 is 3.0 days. The full list with gaps is in codex-reset-history.csv.

The odds, worked out

Treat resets as random arrivals at the recent pace: 20 resets in 68 days is a rate of 0.29 a day. The chance at least one lands within h hours is 1 - e-0.29 × h/24.

Chance another public reset lands within the next N hours, at the Aug to Oct 2026 pace0%25%50%75%100%14%12 h25%24 h44%2 days59%3 days87%7 days
Chance another public reset lands within the next N hours, at the Aug to Oct 2026 pace

Now apply it to a banked reset. Our full Pro $100 windows took a median of about 25 hours to burn. Redeem a banked reset and there is roughly a 26% chance a global reset lands before you finish it. On average it lands halfway through, so the expected loss is about 15% of a window, or $27 at our $175 median. That is the price of redeeming early. Redeeming while your gauge is not yet full is far worse: redeem at 60% and you throw away 40% for certain.

Your situationWhat to do
Gauge under 100%Do not redeem. You lose whatever is left.
Blocked, natural reset in under a dayWait if you can. A global reset in that time is about 25% likely and costs you nothing.
Blocked, natural reset days away, heavy work queuedRedeem, and start the heaviest work right away so the window burns fast.
Blocked, light work onlyHold. A slow window has the most exposure to a global reset.
A model launch or outage was just announcedHold. Launch days have repeatedly come with resets, such as September 5 and September 22.

Check before you redeem

For the full 12-day log behind this, see our 8-window reset log and what the resets were worth. For dated reset history, see the Codex reset log.

Sources

Related


Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →

Back to finopsllm.com