OpenAI Resets Codex Usage Limits and Fixes Eight Waste Issues
OpenAI Codex and ChatGPT business head Tibo announced that usage limits for all paid users of Codex and ChatGPT Work have been reset, stating that the repaired equivalent quota can be used approximately 10% to 50% more depending on usage patterns.
The team reported that they processed thousands of reports overnight. The compression process retains old images in the context, which are large enough to trigger compression again, causing heavy image users' usage to decrease by about 10%. The background memory process inherits stop hooks but cannot end, affecting less than 1% of users, with extreme threads inquiring whether they could stop 15,000 times.
Target instructions sometimes continue running after completion or repeatedly retry damaged tools, with a single instance consuming 15% to 70% of the weekly quota. Some custom automations execute more frequently than set. Smaller models like Luna call stronger assistants when not explicitly requested; orchestration models run sub-agents in fast mode even when not in fast mode.
Old computer history may repeat summary overlapping activities, with some cases consuming one-fifth of the quota in a week. Regular rounds trigger additional rolling task summaries, increasing tokens by about 1%. Some MCP tool results were encoded twice, and tool descriptions were truncated before being fetched. The above issues have been fixed or closed, and architectural changes and duty alerts have been added to prevent recurrence. In-app usage direction display is still under development.
This is not the first reset this year. At the end of June, a reset occurred due to more proactive automatic reviews, increased sub-agents, and excessive retries after backend suggestions failed, along with an additional quota gift; in July, after the launch of ChatGPT Work and new models, another double reset occurred due to high computing power defaults being insufficiently transparent, changes in desktop information architecture, and multi-agent rollbacks. In mid-August, when Codex active users reached 20 million, users were given a self-service accumulation reset and an investigation into converting subscriptions to interface traffic for redistribution was conducted.
In market mechanics, buyers are paid developers coded by weekly quotas, while sellers are laboratories stacking agents, compression, memory, and tool calls on the same bill. The event was driven by reports from users consuming quotas too quickly. Funds still flow in through subscriptions, while idle tokens directly consume gross profit and GPU time. Beneficiaries are paid users whose quotas are taken by background tasks, while the burden falls on the design of agent products lacking billing transparency.
Source: Public Information
ABAB AI Insight
After OpenAI turned the coding assistant into a subscription model, the quota became the steering wheel for product managers. Tibo's repeated resets are not a customer service stance but a way to turn the funnel from "stronger models" back to "completing a week at the same price." Automatic reviews in June, high computing power defaults in July, and 20 million users in August, each time the agent function writes hidden tokens into the backend. Codex needs to serve command lines, desktops, and ChatGPT Work simultaneously, yet the billing unit remains weekly quotas; as the system becomes more complex, idle operations turn into financial loopholes.
The capital path is: first use resets to retain users, then fix compression, memory, and sub-agent loops that can self-replicate, and finally commit to displaying usage direction on the page. The motivation is that if subscriptions feel like "money evaporating in the backend," renewal rates will drop before model evaluations. The strategy is to list faults instead of denying them, using repeatable percentages (10%, one-fifth, 15,000 stop checks) to replace vague "we are investigating." Every segment of wasted computing power is inventory that could have been sold to API customers.
In contrast to Anthropic's less itemized communication on quotas, editors like Cursor writing request counts into package pages, and cloud vendors' billing disputes over idle instances, the coding agent industry is in a period of expansion where functional stacking outpaces measurement systems: sub-agents, MCP, and memory workers are all coming online, while billing semantics still linger at "you clicked once." OpenAI uses public lists to compete for trust rather than lower prices to compete for market share.
The essence is a transfer of pricing power. Users purchase weekly production capacity for task completion, while the company actually deducts tokens from the model's self-circulation. The mechanism is: the more the agent resembles an operating system, the more the background process resembles an uninvited tenant; whoever drives the tenant out of the bill first can increase perceived usage under the same gross profit, thus preventing developers from migrating from subscriptions to pay-as-you-go interfaces or competing editors.
ABAB News · Cognitive Laws
The extra round run in the backend is the real price increase
Quota resets are retention tools, not production increases
The more agents resemble operating systems, the more measurement must resemble electric meters