OpenAI has identified and patched eight bugs in its Codex platform that were causing users to burn through weekly usage limits at an unexpectedly high rate. The fixes follow an investigation prompted by user reports of rapid limit depletion.
Among the issues identified, background processes could enter loops that continued running after tasks had already completed. In at least one documented case, a looping process consumed up to 70% of a user’s weekly allocation on its own. The Computer History feature was found to use up to one-fifth of a user’s limit independently. Subagents were also launching more expensive AI models without explicit user instruction, and errors in Model Context Protocol connections caused the same data to be transmitted multiple times, generating redundant token consumption.
Image handling during compression of long conversations was also found to be miscounting usage, contributing additional unintended drain across extended sessions.
OpenAI estimates that users should now receive approximately 10% to 50% more productive output from the same weekly limits following the fixes, with the range varying depending on the type of workflow. As a result of the issues, OpenAI has fully reset the used limits for all paid Codex and Work subscribers.
The company also said it plans to introduce more detailed usage statistics, giving users clearer visibility into exactly how their token allocations are being spent across different tasks and system processes. No timeline for the usage dashboard update was provided.
