Codex Head Says Good Products Require at Least 34 Days
Thibault "Tibo" Sottiaux, head of OpenAI Codex, stated: "Good products take time, at least 34 days." The post did not specify the product, feature, release date, or specific development milestones.
The mention of "34 days" has sparked speculation in the community about the release date of OpenAI's next-generation model, Astra, but this is merely a third-party interpretation; Sottiaux did not mention Astra, GPT-6, Codex version updates, or any countdown arrangements in the original post.
Sottiaux's recent public communications have mainly focused on managing computing quotas for Codex and ChatGPT Work. OpenAI has reinstated a five-hour usage window limit for ChatGPT Plus users since August 25 to smooth out computational load and prevent users from exhausting their weekly quota in a single long session.
This quota adjustment does not affect Pro, Enterprise, and Edu subscriptions; Sottiaux stated that Pro users will not see the five-hour limit reinstated in the coming months. This move reflects the ongoing demand for reasoning power due to high-intensity proxy programming and long-context workflows in Codex.
Sottiaux previously mentioned that Codex and ChatGPT Work reached 20 million active users in the week of August, and provided paid users with a "banked reset" quota that they could choose to use at their convenience; this user data was disclosed by him, and OpenAI has not released an independent product statistics page.
From a market mechanism perspective, a short internal post from a product lead without product details does not change model supply or OpenAI's revenue but does drive expectations and social discourse around Astra, GPT-6, and Codex upgrades. The real variables affecting users and developers remain whether the model is released, its performance-to-price ratio, API and subscription quotas, context capacity, and reasoning latency; before these disclosures, "34 days" can only be seen as an unverified contextual statement.
Source: Public Information
ABAB AI Insight
The recent historical context of Codex better explains this post: the growth in active users coincides with controversies over reasoning quotas. Sottiaux stated on August 21 that Codex and ChatGPT Work reached 20 million active users and offered a one-time quota reset; days later, the team reinstated the five-hour window limit for Plus users. This indicates that the constraints on product iteration are not just about the software development cycle, but also about how to manage peak loads within limited GPU reasoning capacity after user demand increases.
From a capital perspective, subscription quotas are a tool for tiered pricing of expensive reasoning costs. The usage window limit for Plus users can flatten demand from peaks, preventing a few long tasks from consuming the total weekly quota; the lack of similar limits for Pro, Enterprise, and Edu indicates that higher-paying or contract customers have better computing guarantees. If the model team wants to launch stronger proxy features, they must simultaneously reserve capacity for training, evaluation, safety, and online reasoning, thus affecting the product release date due to infrastructure scheduling.
Historical comparisons can be seen in the early capacity limits of ChatGPT, peak throttling during GPT-4, and tiered usage quotas for Claude and Gemini: cutting-edge models often manage scarce computing power during the release period through subscription tiers, queuing, and rate limits, and then increase quotas as data centers and chip supply expand. The uniqueness of Codex lies in that a single coding proxy task may execute continuously, call tools, read large amounts of code, and perform multiple rounds of reasoning, consuming marginal computing power far more than short Q&A.
This belongs to capital concentration. When the experience of the strongest models is determined by reasoning power, platforms that have GPUs, data center power, cloud scheduling capabilities, and high ARPU customer bases can convert computing scarcity into product tiering and pricing power; even if developers have similar model capabilities, it is difficult to sustain competition in peak service, long task stability, and subscription subsidies. The phrase "good products take time" is closer to "capabilities, reliability, and costs must converge simultaneously," rather than being interpreted as a definitive release date signal.
ABAB News · Cognitive Law
The bottleneck for product launch is often not the model completion, but the available computing power.
Free experiences rely on computing subsidies, while commercial moats depend on quota design.
The stronger the proxy, the more time must be converted into infrastructure.