OpenAI Codex team engineering lead Thibault Sottiaux announced that the company has identified and fixed a bug that caused Codex, its AI-powered coding assistant, to consume usage quotas abnormally fast. All affected accounts have had their quotas reset.
https://twitter.com/thsottiaux/status/2058280452851638313
Sottiaux explained that the issue stemmed from a recent performance optimization. During rollback analysis, the team discovered that cache hit rates dropped sharply when long session contexts were compressed. As the system tried to compress lengthy conversation histories to free up space, the cache failed to hit, forcing Codex to recompute and reprocess the entire context — burning through tokens at an alarming rate.
OpenAI has now removed the problematic optimization code, and the system is back to normal. Since the bug led to unintended token deductions, the company reset Codex usage quotas for all accounts to compensate affected users.