OpenAI Codex Product Head Tibo: Three Defects in Astra Fixed, Midnight Reset Announced
OpenAI Codex Product Head Tibo issued an update to Astra users, stating that the team has verified recent quality feedback with some users, confirmed and fixed multiple defects, and announced a usage reset before midnight local time today.
Old model skills were reported to trigger too frequently. Tibo stated that some skills written for previous models would intervene too often or block the model's self-checking, leading to deviations during task progression and missing self-checks. The team has completed repairs for this triggering path, aiming to return skill calls to a verifiable state rather than forcing old workflows into the new model.
An optional context management experiment has been closed. This experiment may have caused premature stopping of writing or allowed the model to respond to older messages from the current session. Tibo provided a rough estimate: about 4,000 to 5,000 users were affected by this experiment. The experiment was opt-in but has been withdrawn from online pathways and will no longer be a default or optional diversion.
Misconfigured engines have been removed. Tibo stated that some engines were incorrectly configured, causing measurable quality degradation in the long-tail traffic flowing through these engines. The issue did not manifest as a total crash but was concentrated in tail requests: the completion, follow-up consistency, and result verification of the same type of tasks were significantly worse on the wrong engines. The related engines have now been taken offline.
The team has also made several smaller changes. Tibo listed three sensory goals: more consistent follow-up execution, more accurate tracking of the user's latest message, and more complete checks of intermediate results during task progression. He also mentioned that publicly posted failure examples by users and those who directly assisted in troubleshooting were key inputs for quickly locating the issues.
From a market mechanism perspective, this is an event-driven quota replenishment rather than a regular billing cycle rollover. On the buy side are heavy users who have paid and are consuming Astra/Codex quotas; the reset is equivalent to re-adding tokens wasted due to defects back into their accounts. There is no public secondary market clearing on the sell side, and the pressure falls on OpenAI's own computing power and quota pool. Beneficiaries are paid users affected by skill misfires, context experiments, and faulty engines; the burden falls on the product and infrastructure teams that need to complete routing switches, engine takedowns, and full resets before the midnight window. Funds do not leave the platform but are redistributed once between "paid subscriptions - actual usable computing power."
Public interactions show that this update post received about 12,700 likes, around 1,900 replies, and approximately 910,000 views, with a third-party reset tracking site in the community marking it as a high-certainty signal within the midnight window. Tibo also mentioned around the same time that GPT-5.3-Codex-Spark will be retired next week due to declining usage and the availability of stronger models, which is part of the product closure during the Astra transition period.
Source: Public Information
ABAB AI Insight
Tibo has repeatedly handled Codex incidents with the same approach: first acknowledging routing or risk control misfires, then issuing full or targeted resets within hours to a day. In March 2026, a problem caused requests to be misjudged as high network risk and intercepted, with a reset of rate limits occurring immediately after an approximately 8-minute fix; in June of the same year, excessive marking by anti-fraud mechanisms led to some accounts consuming quotas faster than expected, with investigations stating "no widespread damage was observed," yet a usage reset was still issued to all Codex users. Here, the reset is not customer service appeasement but treats paying users as an implicit SLA in an internal gray environment.
The capital path has not exited the OpenAI subscription pool. Before and after the launch of Astra, Tibo promised compensation according to "if paid ChatGPT plans do not receive Astra, a reset will be stored daily," and during the celebration of approximately 25 million active users, a full refresh was given to all paid ChatGPT Work and Codex subscriptions. Money has not been directed towards external model companies or new computing cluster announcements but has instead been released in advance corresponding to the annual/monthly fees collected, allowing users to continue running long tasks, multiple agents, and skills on the Codex harness. The motivation is to lock in the transition window: if the new model drops in quality regarding skills, context, and engine routing, the paid duration will be wasted, and the renewal rate will be damaged before the model evaluation scores.
Similar operations can be compared to GitHub Copilot's early quota compensations, Cursor's usage replenishment after agent crashes, and Anthropic's temporary limit relaxation for Max during the Claude Code failure window. The industry position has shifted from "expanding model catalog" to "controlling runtime": in OpenAI's own Astra introduction, cross-window notes and retrievable old contexts are written as upcoming default experimental capabilities, with 26.6% of Codex users already using skills, and over 10% of users concurrently running three or more agents weekly. At this moment, what is being fixed is not the benchmark score but the controllability of the agent stack on long-tail traffic.
Structural changes belong to the technical replacement layered on top of regulatory lagging behind runtime centralization. The simultaneous occurrence of old skills, opt-in context experiments, and faulty engines indicates that the replacement is happening at the harness layer rather than simply swapping the base model: whoever controls routing, skill triggering, and context compression determines the ratio of tokens that actually convert into verifiable code. Mechanically, long-tail traffic is the cheapest yet hardest to monitor, and configuration drift will first manifest there; using resets to buy time is because the immediate cost of quality incidents under a subscription model is quota waste, not stock price pricing on the same day. Therefore, pricing power shifts from "model name" back to "who can stably complete a multi-hour task."
ABAB News · Cognitive Laws
Gray users are not customers; they are free quality inspection production lines.
Quota resets are implicit warranty clauses of the subscription model.
Long-tail traffic first exposes the system, with averages later declaring victory.