At the TechCrunch Disrupt conference on 26 October 2026, Alexander Embiricos, who leads enterprise products at OpenAI, described the path the dots assistant took. The product did not come out of new technology but out of repackaging agentic tooling built earlier for software engineers into a single continuous conversation with an "entity". The company is also in talks with payment processors, which would open the way for the assistant to spend money.
Key takeaways
- dots was released at OpenAI's developer day in late September 2026, dated 29 September on the company blog
- Available in the ChatGPT desktop and mobile apps to subscribers on the Pro tier or higher — from $100 a month
- A "proactive research" mechanism periodically reviews what the assistant knows about the user and proposes actions on its own
- The name plays on the negative space inside OpenAI's logo, and the dot.com domain is owned by Elon Musk
- The app refuses read-only access to cloud drives, and its permission controls are not granular
Let us picture a situation
An engineer at a drone delivery startup notices a service outage. They open the console to work out what happened — and find an analysis already waiting, which nobody requested.
Their assistant had spotted the outage earlier, analysed it, and come back with information on how it could be resolved. That is the example Embiricos cited at Disrupt as the model use case for dots.
But here is the problem: to notice an outage unprompted, an assistant has to be watching constantly — and to watch, it needs access.
That trade is what the whole product is built on.
What exactly did OpenAI do?
Dots is an assistant that keeps working in the background even when nobody is asking it anything. OpenAI released it at its developer day in late September 2026. It is available in the ChatGPT desktop and mobile apps to subscribers on the Pro tier or higher, from $100 a month. Once calendar, email, computer and other tools are connected, the agent tries to get work done on the user's behalf.
The most interesting part of Embiricos's account is what dots does not contain. Much of the functionality could, in his words, be achieved with the Codex app — the coding tool released at the start of the year — it just takes more work and hand-holding.
As we were working on the product, we realized if we could frame this as, it's an entity and it's a single conversation, not many different conversations, it starts to become intuitive.
Alexander Embiricos, head of enterprise product at OpenAI, at TechCrunch Disrupt.
The same logic earlier produced ChatGPT Work, spun out of the agentic coding tools. The name itself — a play on the negative space inside OpenAI's logo — was meant, Embiricos says, "to feel like just something that could eventually be ubiquitous". The dot.com domain belongs to Elon Musk, which Embiricos admits he found out very late.
How does "proactive research" work?
Asked whether dots collects user data comparable to the profiles built by Meta's Muse, Embiricos said yes, and named the mechanism "proactive research".
| Assistant | What it collects | Business model |
|---|---|---|
| dots (OpenAI) | Periodically reviews what it knows about the user under "proactive research" | Subscription from $100 a month |
| Muse (Meta) | Profiles of family and friends — publicly criticised | Tied to an advertising business |
| Hark | States it collects no data for advertising | No advertising business |
This is a simplified rendering based on a spoken description at the conference, not on product documentation.
According to OpenAI the assistant has the same privacy controls as ChatGPT itself: model training on a user's own conversations can be switched off, and access to resources on the machine can be scoped. Those controls are not granular, though, and the app can be pushy — it refuses, for instance, to accept read-only access to cloud drives.
Embiricos conceded that at launch the company drew criticism from the opposite direction: the defaults were very conservative and the agent asked for permission too often. He said the plan is to build "super granular" controls, while pointing to the model's own intelligence and discretion as the other side of the same coin.
Why not simply use an on-demand assistant?
The usual arrangement today: a user asks a question, the model answers, the session ends. An on-demand agent needs no standing access and builds no profile — which is a strength, not a flaw.
What changes is that some useful work is anticipatory by nature. A calendar conflict, an approaching deadline or a service outage are only valuable if someone spots them in time. A model queried on demand cannot do that, because nobody queries it until the problem is already visible.
The price has two parts. The first is access: to monitor anything, the assistant needs email, calendar and desktop. The second is tokens — a continuously working agent consumes far more of them, which is revenue for OpenAI and the cost of a plan from $100 a month for the user.
There is also a second-order problem that access alone does not solve: timing. Embiricos described his own assistant warning him about a calendar conflict three minutes before the event. The information was correct and still useless.
Is it live yet?
Yes, it is a commercially available product as of late September 2026. OpenAI's blog dates the announcement to 29 September 2026. Access is for Pro-tier and higher subscribers, from $100 a month, in the ChatGPT desktop and mobile apps.
On TechCrunch's testing the anticipatory capability is apparent and often useful but not yet polished — the assistant flags calendar conflicts and upcoming deadlines, not necessarily at the right moment.
Payments are an announcement rather than a feature for now. Embiricos said the company is "deep in lots of conversations with payment processors" but gave no date or scope.
Why does it matter?
The product shows that the edge in this category no longer comes from the model but from the packaging — the same capability sold as one entity rather than a toolkit reaches a different buyer. It also exposes a real boundary: an agent can be handed an inbox, but a bank account is an entirely different decision.
What comes next?
- OpenAI has said it will build granular permission controls — today the app will not even accept read-only access to cloud drives.
- The company is in talks with payment processors, which would be its first case of an assistant spending money. Embiricos gave no date.
- The scale problem on the receiving side remains unsolved: businesses and websites are not equipped to deal with agents, and Embiricos himself said he never intends to have a meeting with someone else's agent.
