OpenAI has reset usage limits for all paid Codex and ChatGPT Work users after resolving a series of technical issues that were causing users to consume their available usage allowances more quickly than expected. According to the company, the fixes are expected to provide between 10 percent and 50 percent more usable capacity, depending on how customers use Codex and its associated features. The improvements follow an internal review of thousands of user reports, during which OpenAI identified multiple issues affecting compaction, memory management, automations, subagents and other background processes. As part of the update, paid users have received a fresh usage reset to restore their allowances while the improvements take effect.
One of the primary issues involved the compaction process, where older images sometimes remained in the conversation context even after they were no longer required. This caused conversations to become unnecessarily large and, in some cases, immediately triggered additional compaction cycles that consumed extra usage. OpenAI stated that resolving this issue alone reduced usage by approximately 10 percent for users who frequently work with images. The company also corrected a memory management issue affecting fewer than one percent of users, where background memory workers continued running because inherited stop hooks prevented them from completing their tasks. In one example shared by OpenAI, a background thread repeatedly checked whether it could stop more than 15,000 times before finally ending. Additional problems involving Goals allowed certain tasks to continue running after reaching their intended completion point or repeatedly retry failed tools, with some cases consuming between 15 percent and 70 percent of a user’s weekly allowance. The company also addressed an issue where custom automations occasionally executed more frequently than users had configured.
OpenAI identified several other factors that contributed to unnecessary usage consumption. Smaller models, including Luna, were sometimes calling more powerful subagents even when users had not requested them. Researchers also found situations where an orchestrating model operating at standard speed instructed its subagents to run in fast mode, resulting in higher usage than intended. Another issue involved the older Computer History system, which repeatedly summarized overlapping activities and, in some cases, consumed as much as 20 percent of a user’s weekly allowance. OpenAI also disabled rolling task summaries that generated unnecessary background requests during normal conversations, accounting for approximately one percent of token usage. Additional fixes addressed situations where MCP tool results were encoded twice and where incomplete tool instructions triggered repeated retrieval requests, both of which contributed to higher than expected resource consumption.
Alongside these fixes, OpenAI announced architectural improvements designed to prevent similar issues from occurring in the future. The company said new monitoring mechanisms will automatically alert its engineering teams if comparable problems are detected again. OpenAI is also developing improved usage reporting that will allow users to see exactly how their ChatGPT Work and Codex allowances are being consumed across different models, tasks and features. By providing greater transparency into usage patterns, the company aims to help users better understand which activities consume the most resources while ensuring that platform efficiency improvements translate into more productive use of their available allowances.
Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.
