Codex Token Usage Optimization and Quota Reset
Codex announced a quota reset for all paid users of Codex and ChatGPT Work, alongside a series of token usage optimizations. The team investigated thousands of user reports and released fixes that increase usage duration by 10% to 50%. Key fixes include: Compaction no longer retains old images, reducing consumption by ~10% for heavy image users. Memory processes no longer get stuck in idle loops due to stop hooks (extreme case: 15,000 checks). Goals now properly stop after completion, preventing wasted retries that previously consumed 15-70% of weekly quota. Automations now respect user-set frequencies. Subagents no longer incorrectly call more expensive models or misapply fast mode. Computer History no longer repeatedly summarizes overlapping activity, which had consumed up to 20% of weekly quota. Rolling task summaries were disabled, saving ~1% tokens. MCP double encoding and truncated instructions were fixed. Architectural safeguards and alerts were added to prevent recurrence. An in-app usage breakdown feature is in development.
Key insights
Quota Reset Accompanies Significant Token Optimization Fixes
Codex reset quotas for all paid users while simultaneously rolling out fixes that increase usage duration by 10-50%. The team conducted a thorough investigation of user reports and addressed multiple bugs that had been silently draining quotas.
Memory and Goals Bugs Caused Severe Quota Drain for Some Users
Memory processes could get stuck in idle loops checking 15,000 times whether they could stop, while Goals could continue running past completion and retry dead tools, consuming 15-70% of a user's weekly quota in extreme cases. Both issues have been fixed.
Subagent Misbehavior Led to Unnecessary Expensive Model Calls
Smaller subagent models were found to call more expensive assistant models without instruction, and scheduling models incorrectly forced subagents into fast mode. These behaviors have been corrected to prevent unintended quota consumption.
Computer History Bug Alone Could Consume 20% of Weekly Quota
An old implementation caused repeated summarization of overlapping activity records, eating up to one-fifth of a user's weekly quota. This has been fixed, along with other inefficiencies like rolling task summaries that added ~1% token overhead.
Architectural Changes Aim to Prevent Future Regressions
Codex implemented low-level architectural adjustments and alerting systems to prevent these issues from recurring. Additionally, an upcoming in-app feature will show users exactly where their quota is spent, increasing transparency.