OpenAI Issues Codex Reset Cards, Users Report Halved Quota
During the phased rollout of GPT-6 Astra to Plus, Pro, and Business users, OpenAI issued Codex quota reset cards to paid users who received the model later. The help center clarified that the reset card is a one-time usage reset, stored in the account for later use or expiration, not purchased credits, and does not permanently increase the plan limit; existing reputable Plus, Pro, and Business users received one card each on September 3 and 4, while new accounts are eligible based on the Pacific Time cutoff, with credits possibly arriving a few hours late.
Some users compared their usage records before and after the reset, consuming about 14% of the weekly quota. Before the reset, they used approximately 114 million input tokens and received 738 model responses; after the reset, they used about 63.6 million tokens and received 489 responses. By extrapolating the same percentage, they estimated that the available amount corresponding to the reset card was only half of the normal weekly quota. The comparison is not clean: both segments primarily ran Astra, but the inference settings and agent usage were different. Codex head Thibault Sottiaux later denied any difference in available quotas before and after the reset.
Astra itself is more expensive and consumes resources faster. According to the official rate card, GPT-6 Astra is charged 250 credits per million input tokens, with 25 for cached input and 1250 for output; GPT-5.6 Sol corresponds to 100, 10, and 500 respectively. Astra's Fast mode is charged at 2.5 times the standard rate. In the Plus tier, during a roughly 5-hour window, Astra local messages are about 5 to 45, while Sol messages are about 10 to 100. Some users reported that a single Astra prompt could deplete the entire 5-hour quota. Codex and ChatGPT Work share the agent usage pool, and heavy loads can simultaneously consume resources from both sides.
The reset card has become a tool for compensating releases. When Astra's launch was delayed, the official issued a card for each day that paid plans did not utilize Astra; after the early release, they issued full quota reset cards to Plus, Pro, and Business users again. The stored reset differs from an immediate purchase reset: the latter immediately activates the 5-hour and weekly windows and resets the 7-day clock from the next Work or Codex request, and cannot be saved for later use. The specific weekly quota numbers have never been disclosed by the official; the pricing page only states that there may be additional weekly limits.
This is not the first time there has been a dispute over usage. At the end of June, the official stated after an investigation that there was no single master switch, but rather an automated review process that was more proactive, with sub-agents being triggered additionally, and backend suggestions for repeated generation or retries after failures being too frequent, causing reports to mistakenly attribute automatic reviews to old models, while failed requests still showed as rounds. At that time, the approach was to fully reset and then issue a stored card, while rolling back backend consumption. In August, some users again reported rapid quota depletion, and the official stated they had not seen any anomalies, while marking some traffic as behavior of converting subscriptions to API for redistribution.
In market mechanisms, what is consumed is the weekly quota and 5-hour window within the paid plan, while what is delayed in delivery is Astra's computing power. The event was driven by the phased rollout of the model and the overlap of compensation card issuance. The pricing of credits reflects stronger models at higher prices, and the same percentage progress bar corresponds to fewer tokens on Astra. Benefiting are heavy users who received Astra early and can stack reset cards on high inference settings; pressured are ordinary subscribers who judge whether the "card is sufficient" based on counts and percentages. The official denies that the quota has been cut in half, while users find it difficult to compare due to task structure. The dispute lies in the measurement criteria, not in whether cards were issued.
Source: Public Information
ABAB AI Insight
OpenAI changed the throttling from "per message" to token credits, leading users to see percentages, while the company accounts for model pricing multiplied by tokens. Astra's rate is 2.5 times that of Sol, and Fast mode multiplies it again by 2.5, meaning the same weekly progress bar will inevitably move faster after switching models. The reset card does not change the rate card; it merely resets the window back to 100%. Understanding "reset back to 100%" as "able to produce the same number of tokens as last week" does not hold when the model has already switched. Thibault Sottiaux has repeatedly used card issuance to calm release incidents, with active users surpassing 7 million in July, 20 million in August, and 25 million, all receiving cards, indicating that compensation has become a product.
The capital path is the time transfer of computing quotas. Phased rollouts mean that GPUs are initially given to a portion of paid accounts, while others wait using stored resets. Credits can be purchased separately and automatically replenished; reset cards expire in 30 days, while purchased resets take effect immediately and reset the 7-day clock. The company uses card issuance to reduce churn and manage public opinion, while using the rate card to recover Astra's marginal costs. The automated review and sub-agent quota theft in June indicate that there are tasks in the progress bar that users have not clicked on; users compare tokens and response counts, but lack the breakdown of inference settings, tool calls, and backend tasks.
Comparisons need to be specific. APIs are priced transparently by tokens, while subscriptions hide the same measurement in weekly percentages, with disputes arising each time from "invisible work." Anthropic and Google document long contexts and tool calls in their unit prices, while OpenAI combines Codex and ChatGPT Work into the same pool, allowing a single agent link to simultaneously hit quotas for two products. The industry is at a stage where "model releases rely on compensation cards, and measurements depend on post-hoc explanations," rather than quietly halving weekly quotas—at least the official written definition remains to restore normal limits, not to issue half-value vouchers.
Structural judgments belong to the transfer of pricing power. The mechanism is: when the pricing unit changes from message counts to model-weighted tokens, the progress bar is no longer a neutral measure but becomes a display of the rate card. Users extrapolating absolute amounts from the same percentage before and after the reset equate to measuring the new model with the old model's ruler. Whoever defines "how many tokens equal one reset" defines the true capacity of the subscription. Card issuance soothes waiting but does not rewrite the unit price.
ABAB News · Cognitive Law
Progress bar unchanged, unit price changes means a disguised reduction in quantity
Compensation vouchers reset the window, but do not reset the rate
Incomparable comparisons, the dispute is over measurement rights, not emotions