Why Codex burns a weekly limit in a day while the agent waits for the tide (relux.works)

A recent analysis revealed an input-token leak in Codex models, sparking frustration among users as their $200 weekly subscription often runs out in a day. The issue stems from how the "goal mode" in Codex processes tasks; it inadvertently accumulates high token usage due to inefficient waiting mechanisms. Specifically, during a session, the model polled for updates every 0.03 seconds without pausing effectively, leading to an astonishing consumption of input tokens—over half of the 59.2 billion used in the summer came from just 52 sessions. This discovery holds significant implications for the AI/ML community, highlighting the importance of optimized waiting strategies in model architectures. The findings suggest potential fixes, including the integration of a “sleep” function that could reduce unnecessary token consumption during idle periods. As the developers work on solutions that could streamline operations and prevent excessive costs, this incident underscores the need for transparent billing practices and effective resource management within AI systems, raising questions about how token usage is calculated and the impact of caching on subscription limits.