OpenAI Codex Lead Tibo Pledges Daily Improvements or Quota Resets for 28 Days

The promise Tibo Sottiaux made on X can be checked day by day: over the next 28 days, either ship one improvement clearly useful to most Codex and Work users every day, or perform a full quota reset. The structure of this sentence itself is worth noting—it doesn't promise "how many features will be released," but hands the judgment of "is it useful" over to users.

What supports this promise is his description of his own authority. Tibo says that whenever it's necessary and he deems it appropriate, he can just press the quota reset button directly, without letting this go through layers of approval. When asked how many more resets to expect, his answer was "depends on how many times we break things," adding that there will be resets when problems arise, and resets when there's something worth celebrating.

This self-deprecation isn't without basis. He admits that on his third day at OpenAI, he once took down the production environment. Putting this experience together with "depends on how many times we break things," the quota reset in his telling looks more like a compensation mechanism for mistakes than a purely marketing move.

More noteworthy than the quota is Tibo's judgment about the current AI experience. He says outright that he too gets exhausted by options like "which model to pick, how high to set reasoning effort, whether to use multi-agent mode or Ultra mode," describing it as "you practically need a 'model selector PhD.'" He hopes to shed these burdens as soon as possible, letting the app itself gradually fade into the background.

This pursuit of a simple experience also comes from changes in his own work. He says that nowadays a large amount of the code needed to analyze business trends, research the next feature, and check the performance of previously shipped products is written by Codex, and the only handwritten code left is when he relaxes on weekends doing a few LeetCode problems.

It's worth noting that he doesn't single-mindedly pursue running more agents at once. He says that as agents get faster, he has actually reduced the number he runs simultaneously, rediscovering his work flow, and says he no longer misses the old state. A cautious inference can be made here: if even the person most familiar with this toolset is actively shrinking his concurrency scale, then "stacking more agents" is very likely not the direction he sees for the product.

Tibo believes the next phase of product design needs to account for three changes: most operations on the internet will be performed by agents; models will rapidly become cheaper and faster; and interaction across different modalities will become smoother.

These three premises explain the logic behind his skepticism toward some outside products: if you seriously imagine these capabilities being ten times stronger a year from now, developers would build in different ways. In other words, an architecture that looks reasonable today may just be a compromise made for today's capability ceiling.

From this one can also understand why he thinks manually building loops and repeatedly tuning agent workflows may just be a transitional phase. It should be noted that this is his personal judgment, not an established fact.

Tibo confirms the long-term direction is to make Dots the primary entry point for human-AI interaction, which then launches tools or services like Codex and ChatGPT Work. But he also leaves a door open: whether or not it ultimately is Dots, the fundamental goal is to free people from the constraints of the technology itself. This means the direction itself remains to be validated, and Dots is merely the currently chosen vehicle.

On the integration path, he says he will integrate Codex and ChatGPT, merging the chat mode and Work mode experiences to reduce complexity, and ultimately bring Dots' capabilities directly into chat mode. He mentions ChatGPT's user scale of roughly 1.2 billion, but also states clearly that by the time Dots' capabilities are brought into chat mode, the user count will have to be seen—this number is not a fixed value.

Dots' product form is described quite restrainedly: no model selector, and no need for various configurations; users just need to talk to it, and the only thing to configure is which channels they want to communicate with it through. For now, one user-configurable main Dot is launched first, and soon users can add a second, a third, or any number of Dots, forming a virtual team and assigning specific roles.

What supports this vision is the research accumulation he describes: at the research level, OpenAI has been conducting research on persistent, long-horizon tasks for over two years, and has studied memory systems capable of maintaining coherence for roughly the same amount of time. He says Dots was chosen to launch on Astra because it is OpenAI's safest model and the one most aligned with its alignment goals. These are all his own statements, with no first-party independent corroboration seen in the sources.

Tibo believes what's underrated is the ecosystem. He says OpenAI has signed 16 partners and turned this certification mechanism into a formally supported mechanism. He traces the starting point of this to conversations with the developers of Pi and OpenCode—users should rightly be able to log in with Codex and use their own quota.

A more specific layer is plugins. He says OpenAI has designed a revenue-sharing mechanism, paying compensation and giving a portion of revenue share for popular, high-usage plugins, but this mechanism was not mentioned in the keynote.

The recommendation mechanism is tied to retention: the system looks at user retention data and plugin quality, then decides whether to recommend it to users in conversation; if a plugin isn't good enough, the system stops recommending it. He says this mechanism will continue to be adjusted, and the specific rules and effects have not yet been announced.

How well the 28-day promise is fulfilled depends on "how many times we break things," and this cannot be determined in advance. What can be determined is the standard of judgment: whether the product gets better cannot be judged only by the number of features shipped, but also by whether users truly benefit in their daily work.

Tibo also admits that the products launched this time are not perfect, and that he will learn from opening them to all Pro users. Dots currently only opens one main Dot, and multi-Dot capability has not yet been launched.

What needs to be distinguished is the level of evidence: the 28-day promise itself is a checkable public statement; while statements about quota reset autonomy, research duration, number of partners, and model safety all come from Tibo himself, and are source claims rather than independently confirmed facts.