OpenAI Core Product Head Tibo Says Millions of Dots Will Be Online in Days
Tibo Sottiaux: Millions of Dots Will Go Live in Days
OpenAI core product head Tibo Sottiaux stated that millions of dots will be online simultaneously in a few days, processing various tasks in parallel among a large and diverse user base. The team will first observe how each person uses their main dot before opening up to form entire dot teams.
Dots are designed to run around the clock, capable of continuous learning based on feedback, equipped with independent computers and browsers, and can connect to over 4,000 applications within the ecosystem. The driving model is Astra, with the main dot included in the Pro package and not consuming regular quotas, allowing users to call it directly while they work. Coverage includes Pro, Business Premium, Enterprise, and Pro 100, with the official statement indicating plans to continue expanding access.
Sottiaux mentioned that he experienced a significant improvement after using it for 2 to 3 days, provided he continuously taught it his personal preferences and ongoing thoughts. He assessed that the system learns quickly and can independently handle unexpectedly difficult tasks while collecting feedback on "which experiences are magical and which are not enough."
He clarified that the main dot is included in the package and can be online 24 hours a day; if a dot is tasked with creating Codex tasks, those tasks will still deduct from the quota as per original rules; however, direct work done by the dot does not consume additional package usage. Future speed increases and bandwidth expansions will be paid options, but basic capabilities will remain tied to existing packages. The new Codex Cloud will launch simultaneously, with the Agents API entering preview and supporting computer operations, using the same technology that drives cloud-based intelligent agents, including dots.
ChatGPT Space will also be released concurrently, allowing users to collaboratively edit real-time visual notes, to-do lists, product plans, and meeting minutes with others and dots in the same space, with automatic background synchronization. The developer side is described as taking AGENTS.md a step further. Community notes indicate that dots will be rolled out in batches according to qualified markets, with Pro not initially available in the European Economic Area, the UK, and Switzerland, meaning not all paying users will have immediate access.
In terms of market mechanism, this is an event-driven supply expansion: OpenAI exchanges permanently resident intelligent agents included in packages for real task data and preference training, first deploying millions of main dots, then deciding when to open up multi-agent formations. Demand comes from already paying Pro and enterprise seats, while supply consists of computing power, cloud environments, and application connection points. Beneficiaries are paying users who can embed preferences into continuously online agents, as well as OpenAI, which consolidates ChatGPT and Codex into a single entry point; those under pressure include pay-per-use coding agents, assistant products lacking independent computing environments, and users excluded from the initial market rollout.
Source: Public Information
ABAB AI Insight
Sottiaux previously worked on machine learning infrastructure at Google Maps and DeepMind, and after joining OpenAI in 2024, he transformed the internal ASWE (Autonomous Software Engineer) into Codex, increasing weekly active users from millions to tens of millions by repeatedly resetting usage quotas. After the product line was downsized in May 2026, he took over core products and platforms, reporting to Greg Brockman, covering ChatGPT, Codex, and API, shifting the task from serving developers to transforming a conversational product with nearly a billion weekly active users into a delegable personal agent. Dots represent the next step in this path: not just another programming window, but a permanent machine for each paying account that can browse the internet and run tasks independently.
The capital path is clear. Astra, independent computers, browsers, and connections to 4,000 applications change interactions previously billed by token into a 24-hour occupancy included in the package. By first providing the main dot without deducting from the quota, it uses computing power subsidies to exchange for behavioral data; speeding up and increasing bandwidth will be reserved for paid options, turning "learning your preferences" into incremental ARPU after that. The simultaneous preview of Codex Cloud and Agents API means selling the same orchestration that runs millions of cloud instances internally to the outside, with dots being both a product and the largest internal testing cluster for this orchestration.
This approach is akin to a second attack after the Operator's failure, and is close to the route taken by Devin and Lovable to make "a computer's agent" a delivery unit, but OpenAI has chosen to tie it to existing Pro seats rather than selling seats separately. The industry phase has shifted from "able to write code" to "able to take over tasks," with the main dot leading the way and team functionalities following, indicating that orchestration and permissions are not yet ready for users to deploy multiple agents at once. In comparative terms, whoever first turns preference memory into a switching cost will be closer to making assistants an account asset.
Structurally, this represents a combination of technological substitution and a transfer of pricing power. The substitution is not for single completions, but for the time users spend monitoring the screen to advance tasks; pricing power shifts from "charging per inquiry" to "whether there is a machine that can work permanently in the seat." The mechanism is: first use free occupancy to integrate millions of agents into real workflows, use 2 to 3 days of preference input to increase retention, and then make multi-agent and higher bandwidth the next tier of charges. Whoever controls this always-online machine will control subsequent task flows, application distribution, and seat renewals.
ABAB News · Cognitive Laws
First provide a machine that never goes offline, then charge for speed
After preferences are taught in, switching costs only start accruing then
Team functionalities will always lag behind individual deployments because orchestration is more expensive than models