Why Codex Reset Your Quota: 8 Hidden Bugs Fixed to Boost Usage Efficiency

On Saturday night the OpenAI Codex team announced that usage for all paid Codex and ChatGPT Work users has been reset; Codex is a programming agent that allocates a weekly token quota to each paid user.

The reset is not a temporary perk. After receiving many “quota not enough” complaints, the team examined thousands of reports and overnight patched eight bugs that were covertly consuming tokens. Tibo said the fixes can increase effective usage by roughly 10%‑50% depending on how the service is used.

Context Compaction : The old logic kept previous images during compression, leaving the context size large and sometimes triggering another compression pass. After the fix, heavy‑image users see token usage drop by about 10%.

Backend Memory : Backend memory tasks inherited Stop hooks and kept running even when a hook disallowed continuation. Affected users are under 1%, but the tail is severe; an extreme case involved a thread checking 15,000 times whether it could stop.

Goals : The /goal command sometimes failed to stop or caused the model to repeatedly retry a broken tool, burning 15%‑70% of a weekly quota in isolated incidents.

Automations : Some custom schedules ran more frequently than their configured intervals.

Subagents : Small models (e.g., Luna) occasionally invoked stronger sub‑agents without the /fast mode, and the main model could also launch sub‑agents in /fast mode unintentionally.

Computer History : The old implementation repeatedly summarized overlapping activities, causing certain users to lose about one‑fifth of their weekly quota.

Rolling Task Summaries : Regular dialogue turns triggered extra backend requests, adding roughly 1% token overhead per turn; this was disabled.

MCP (Model Context Protocol) : Some tool results were encoded twice, and truncated tool commands were fetched again.

These bugs share the trait of being hidden in the details of model‑client collaboration and being amplified over time. The team also made architectural adjustments to prevent regressions, set up automatic alerts for future occurrences, and will display token flow directly inside the application so users no longer have to guess.

Some users remain dissatisfied: one noted that the Computer History feature alone can consume 20% of the weekly quota, and others complained that the reset window of five hours is too short.

Tibo concluded with a friendly “Enjoy your weekend.”