OpenAI fixes Codex long-session usage accounting

• Long sessions with images and repeated compaction, high p95+ Computer History use, and automatic conversation-title generation were consuming more Codex usage than intended, according to reach_vb's update.
• A separate investigation found worse cache hit rates for some users, which @thsottiaux said could also make weekly usage drain faster thsottiaux's cache update.
• OpenAI issued a reset alongside the fixes, and @thsottiaux later said the reset had propagated and the first usage fixes had landed thsottiaux's Sunday update.
• Reset delivery was uneven at first: one customer reported a Business account and a Plus account with an active task missed it HCSolakoglu's report, before @thsottiaux said the Business rollout was still propagating thsottiaux's business-account reply.

Codex’s current rate card says its pricing moved from per-message accounting to API-token usage in April. @thsottiaux’s full incident thread separates the three identified drains from the cache diagnosis, while an OpenAI Developer Community discussion asks whether subscribers could see cache statistics directly.

Before the bug-specific reset, @thsottiaux announced that Codex had passed 20 million active users and said every Codex and ChatGPT Work user would receive a banked reset, redeemable at their own discretion. The same announcement said the company was investigating reports of accelerated draining despite not seeing an abnormal pattern overall.

@thsottiaux’s incident thread named three sources of unintended usage:

1. Images in long sessions that have gone through multiple compactions.
2. High p95+ usage from Computer History.
3. A feature that generates conversation titles.

A follow-up from reach_vb said fixes for those paths and a reset were scheduled for 2 p.m. PT.

The root-cause list did not close the investigation. @thsottiaux later said some users had worse cache hit rates than the stable state of prior weeks, a condition that could explain faster usage depletion.

That distinction fits two reports that did not neatly map to image-heavy or Computer History workloads: vikhyatk said the same two agents began consuming usage twice as fast despite neither factor, while nptacek said 37% of weekly usage disappeared in about 13 minutes. An earlier community summary by aibuilderclub_ had already pointed to lower cache hit rates.

@thsottiaux said the repaired reset had propagated on Sunday and that users should feel a positive difference. When asked about Business accounts, the account replied that rollout was still in progress but should finish within five minutes.

The rollout had visible gaps before that confirmation. bridgemindai reported that the promised reset was absent, then @thsottiaux set an 8 p.m. PT deadline thsottiaux's deadline and subsequently declared it delivered thsottiaux's confirmation. A reach_vb reply also offered to investigate an individual case through a support ID reach_vb's support reply.