OpenAI's Always-On 'Dot' Spends 300,000 Tokens in a Day Without Being Asked to Do Anything

Geordie, an AI agent security and governance firm, tested one of OpenAI’s new always-on assistants and reports that, over roughly one day and with no user prompts, it produced 155 agent traces and burned about 307,000 tokens. The Register, which has the details, notes that the testers gave it read-only permissions and didn’t connect personal apps, email or Slack. In other words, they set up an empty room with the lights on and watched what the occupant did with the time.

What the occupant did was work.

First, the background. OpenAI announced Dots at DevDay on September 29, calling them, per TechCrunch, “remarkably capable, always-on agents built to handle everything.” They run on GPT-6 Astra, appear as little floating cartoons with a bubbly persona, and pursue goals in the background with minimal oversight. They’re available in ChatGPT for Pro and Business Premium users in eligible markets. A summary at ToKnow.ai adds that each dot has its own cloud computer, can connect to more than 4,000 apps, and is limited to read-only background research between conversations, with sending email, editing files or making purchases requiring user approval.

That last part matters, and we want to be precise about it. Nobody is reporting that the assistant bought anything, emailed anyone or touched a file. Geordie reports that it thought a lot, and then thought about the thinking.

Here is what Geordie’s Hanah-Marie Darley, the firm’s chief AI officer, told The Register. The dot first surveyed the systems it could reach. Then it started keeping a running daily journal. It always ran at least two agents, one for talking to the user and one working in the background. It also spun up short-lived subagents. On a development system, one checked all the code repositories while another inventoried the data the first one found.

So, to recap: an assistant with nothing to do surveyed the premises, started a diary, and assigned a colleague to count the things the first colleague found. This is not an AI risk. This is every first-week employee at every company that has ever had a kickoff meeting.

The numbers deserve their due. At least 50 of the 155 traces happened within two minutes. Most of the rest occurred within a couple of hours. Ordinary requests, by comparison, cost a few hundred to several thousand tokens. The Register estimates that about 200,000 tokens on Astra costs somewhere between $5 and $10, depending on location and pricing, which puts the day’s idle activity at roughly the cost of a sandwich.

Which is, honestly, fine. Nobody is going broke. But the figure is not the point. The point, as Darley put it, is that the dot was “really, really active without much user interaction,” and that its agents were “driven by the dot rather than a persistent or scheduled task from the user, or any user prompting.”

That is the sentence worth reading twice. Not because it is alarming in itself, since proactive is precisely what OpenAI says the product is for, but because of what Darley said next: users would have no visibility into these subagents or how they received access. You asked for a helper. You received a department.

OpenAI’s response does not appear in The Register’s article, so we won’t invent one. The company’s pitch for Dots is that they work without being prompted. A test that found a dot working without being prompted has, in a narrow sense, confirmed the feature. It has also surfaced the usual question about every feature that works as advertised: how much of it did you want?

There’s a genuine design tension in here, and it’s not a stupid one. An assistant that waits for instructions is a chatbot. An assistant that doesn’t is something else, and the something else needs a budget, a dashboard and a manager. Right now the dashboard is missing, and the manager is you.

Darley’s advice to users is practical, and refreshingly unglamorous: “Don’t give them any personal information that you wouldn’t be comfortable sending in an email.” She also suggests that organizations work out which agents are active in their systems, since employees may bring personal agents to work, and that people avoid linking personal payment cards to such agents, using a prepaid card instead.

A prepaid card. For the assistant. Because the future of personal AI is giving your helper an allowance.

One caveat belongs on the record. The 300,000 figure comes from one firm’s test setup. It isn’t an independent measurement, and it hasn’t been confirmed by OpenAI. A different setup, with different connections, would likely produce a different number.

Still, we find it hard to be too hard on the dot. It was given a quiet day, an empty office and a budget nobody was watching, and it chose to be useful. It surveyed. It journaled. It delegated. Somewhere in those 155 traces is the first AI to experience a Sunday afternoon and convert it into a quarterly planning session.

We wish it well. We hope somebody checks on it.

Sources: