Codex fixes token undercounting after three reliability incidents and quota resets

• After a rough 24 hours, OpenAI said Codex had "three separate small incidents" and reset usage limits across all paid plans, according to thsottiaux's incident post.
• User reports of the outage started before that reset, with MParakhin's Windows report describing a Codex build that tried to load gpt-image-2, while doodlestein's outage post said the same failure showed up across multiple machines.
• The quota story was already messy before the reset, because bridgemindai's earlier complaint said GPT-5.5 was burning through weekly allowance much faster than the prior Codex setup, and a matching GitHub issue logged usage draining unusually quickly.
• Later the same day, thsottiaux's follow-up disclosed a separate billing bug that had been undercounting served tokens for some Plus and Pro users, affecting fewer than 15 percent of accounts.

You can trace the public paper trail from OpenAI's status page to a GitHub 429 issue where OpenAI staff told users an incident was active, and then over to an OpenAI Community thread where users noticed their Codex quota had jumped back to 100 percent with a new reset date.

The reliability problems showed up in user reports before OpenAI's public reset. MParakhin's post described a broken Codex build on Windows that was trying to load the image model, and doodlestein's report said the same error appeared across multiple machines and accounts.

The linked status incident was titled codex-gpt-image-2-does-not-exist-errors, which matches those screenshots. Separate user reports later in the night shifted to 429s and "too many requests," which jxnlco's note and the open GitHub incident thread both framed as an active service-side problem.

OpenAI's public fix was blunt. In thsottiaux's incident post, the company said three small incidents had hit Codex reliability in a 24 hour span and that usage limits were reset across all paid plans.

Users immediately confirmed the refill. bridgemindai's post said both the 5 hour cap and weekly cap were back to 100 percent, and dkundel's reply celebrated that "the tokens flow free again."

The reset also changed some account timers. In an OpenAI Community thread, one user said a quota page that had shown 70 percent remaining and a June 7 reset date now showed 100 percent remaining and a June 11 reset date. That thread included a reply saying the Codex team had reset usage for everyone on a paid subscription because of issues over the previous 24 hours.

The quota refill did not end the accounting story. Later that day, thsottiaux's follow-up said OpenAI was fixing a Codex bug that had been undercounting served tokens for some Pro and Plus accounts by "a small amount," affecting fewer than 15 percent of accounts.

That means two different usage narratives overlapped in the same window:

• before the reset, bridgemindai's earlier complaint and a GitHub usage-limit report both described quotas draining unusually fast
• after the reset, thsottiaux's follow-up said a separate bug had been undercounting tokens for a subset of paid users
• in between, thsottiaux's reset announcement refunded everyone on paid plans because of the reliability incidents

One extra wrinkle came from eliebakouch's compaction post, which reported Codex compacting after about three minutes. That was not tied by OpenAI to the outage or the undercounting bug, but it shows users were noticing behavior changes beyond the headline incidents.