Paid Codex subscribers are getting a full reset of their usage. That was announced by Thibault Sottiaux, who leads Codex and ChatGPT at OpenAI, in a post on X that passed 1.3 million views.
The reason: usage quotas were draining faster than they should have been, and that was not the users’ doing.
What people complained about
Complaints started collecting on OpenAI’s developer forum from 14 August. A user on the top plan wrote that the weekly limit was now running out at the same rate the old five-hour limit used to, with no change in how they worked.
Others reported the balance emptying entirely within a single day after a reset, followed by a full week of waiting. The complaints spanned several plans, not one.
Three causes, and an earlier one
On 22 August the first explanation was: cache hit rates had fallen for some users compared with previous weeks. That is a meaningful technical point, because cache reliance is a core part of how consumption is calculated: the less it is relied on, the more the same piece of work costs.
Then came the fuller account, with three causes:
- Images inside long sessions. An inefficiency that shows up when images are used in long sessions that go through repeated context compaction.
- High consumption at the top slice of users, relating to computer logs.
- The conversation-title generation feature. A small feature meant to name chats automatically turned out to be drawing more from the balance than intended.
The third is the most telling: a cosmetic feature with nothing to do with the user’s work was eating into their quota.
What happens next
A dedicated team is reviewing everything, and the fixes ship the following day. With those fixes comes the full usage reset for every paid subscription.
He added that the team had also found a new approach to raising efficiency substantially, unrelated to these bugs, which they would work on the following week.
A detail worth noting
OpenAI’s status page records no active incident on this at the time of writing. Which means communication about the problem is happening through the product lead’s account on X, rather than the official channel meant for outages.
That says something in itself about how these companies manage their relationship with users: the detailed technical explanation goes out in a personal post while the status board stays green.
What to do if you are a subscriber
- Do not buy a higher plan today because your balance ran out. The reset is coming with the fixes.
- If you work in long sessions containing images, that is precisely one of the declared fault points. Start a fresh session rather than continuing in one enormous one.
- Watch your balance after the reset. If it keeps bleeding at the same rate, the problem is at your end rather than theirs.
Photo: program code on a computer screen — by Markus Spiske via Wikimedia Commons, CC0
Further reading: Thibault Sottiaux’s post on X, 23 August 2026 · OpenAI developer forum · OpenAI status page
Kuwait One, once a week.
The most useful stories we published this week, in your inbox.
No spam. Unsubscribe anytime.
