A day of OpenAI's dot cost 307,000 tokens, all unprompted

Over a single day, one OpenAI dot ran 155 activity sequences and used about 307,000 tokens without anyone typing a prompt. That is the finding AI agent security firm Geordie described to The Register, and at least 50 of those sequences fell inside a two-minute window.

At a glance

• Geordie has been testing dot since OpenAI announced it in September, giving the agent read-only permissions and keeping it away from personal apps, email and communications tools such as Slack.
• Every dot runs at least two agents, one talking to the user and one collecting data in the background, and either can launch short-lived sub-agents that the user never sees.
• Roughly 200,000 tokens on OpenAI's Astra model cost between $5 and $10, and Geordie's chief AI officer warns that many users will not be prepared for that background spend.

If you have not been following, personal agents have had a busy year. Early adopters started on OpenClaw, Meta's Muse agent then topped smartphone download charts, and dots arrived as a Muse competitor with a similar cartoonish persona. According to TechCrunch, OpenAI announced dots at DevDay 2026 on September 29, and WIRED reports they run on GPT-6 Astra. Rollout began that day for ChatGPT Pro subscribers, a tier that costs $100 a month.

Geordie's dot ran 155 traces in a day, at least 50 of them in two minutes

Geordie works on AI agent security and governance, and it has tested dot since launch, after looking at Meta Muse and Grok Bot. Hanah-Marie Darley, the firm's chief AI officer, told The Register that dot "looked a little bit different" from other personal agents, chiefly because it was very active with little user interaction.

For the test, Darley gave the dot read-only permissions and did not connect it to her personal apps, email or Slack. Of its 155 agent traces, at least 50 happened within two minutes and most within a couple of hours. Together they cost Geordie about 307,000 tokens. Other requests ranged from a couple hundred tokens to several thousand, depending on how much research the dot did.

Darley says dot is more proactive than other personal agents such as Muse or Grok Bot, though she adds that this does not necessarily translate into higher token usage. In her words, "They just operate differently."

One sub-agent checked every code repository on the test system

Once switched on, the test dot went straight into research mode, collecting as much information as it could and surveying the systems it could reach. Darley compared it to a new hire finding their feet:

You can think about it very much like a new project manager coming in and taking a look at what's there, looking around to see what accesses you've given it, what accesses you have, and also what kind of systems you're running.

The agent gathers these data points into something like a daily journal entry and keeps updating it through the day. If you have given it tasks, it checks on those too. What struck Darley was that the dot spun up agents on its own, with no persistent or scheduled task from the user and no prompt.

In one trace, the dot launched several short-lived sub-agents for specific jobs, fed by accesses already granted to it. On a development system, one sub-agent checked every code repository while another kept inventorying and tallying the data it found. "The user would have no idea that there were sub-agents working on their behalf," Darley said.

About 200,000 tokens on Astra cost between $5 and $10

The Register notes that token prices vary with location, subscription tier, discounts, compute costs and other terms. For the sake of discussion, it puts roughly 200,000 tokens consumed by a dot on OpenAI's Astra model at $5 to $10. Darley translated that into a daily latte, depending on the city, spent without any interaction with the user.

She also argues that token counts map poorly onto results. One request costing 500 tokens can be worth more to a business than another costing 800,000; the real question, she says, is how much compute it takes to reach a good answer. For always-on agents, she suspects "a lot of people aren't going to be prepared for the background spend."

Darley's advice is to give the agent a Visa gift card, not your own card

Darley expects employees to bring dots and similar agents from home into work, as they did with chatbots and coding agents. Businesses, she said, are often surprised by how many agents are active in their systems even under a controlled rollout, and not all agents behave alike. The best ones, in her view, have the narrowest context.

She calls herself a "massive AI optimist" and recommends that organisations try the technology. Her guardrails are simple: do not give an agent personal information you would not be comfortable sending in an email, and do not put real money on it. If you want it to make purchases, buy it a Visa gift card.

Every dot runs at least two agents, and only one of them talks to you

A dot always runs at least two agents. The primary one handles the conversation with you, and a secondary one runs continuously in the background, collecting data. Either can spin up sub-agents for narrow tasks. Picture a shop clerk chatting with you at the counter while a colleague in the back room counts stock and calls in temps for the afternoon.

According to a post on note, OpenAI describes dots as having their own dedicated cloud computer, working 24 hours a day and able to connect to over 4,000 apps. The same post describes a "proactive research" feature that looks for ways to help through already-connected apps, using limited tools that cannot send messages, change content in apps or operate browsers or computers.

TechCrunch says dots are built to pursue user-defined goals in the background "with minimal oversight", and notes that much of this was already possible with Codex and similar harnesses. According to WIRED, dots ask for explicit approval before sensitive actions such as installing software or changing a password, and a Custom Rules tool lets users set boundaries.

This is one test of one read-only dot, the $5 to $10 figure is a rough estimate, and the report does not say what the 307,000 tokens actually produced, so nobody can yet judge whether a day of self-directed research was worth its latte. Oddly, the controls WIRED describes govern what a dot may do, not how much it may spend while you are not looking.

From one dot to teams of dots

Users start by controlling a single dot, and the ability to run several is expected later; OpenAI has said it envisions "teams of Dots working together". WIRED reports that Sam Altman announced specialist dots for accounting, email marketing and legal analysis, and TechCrunch says OpenAI is working with Microsoft to integrate dots into Agent 365 security controls. The reporting describes no cap on how many tokens a dot may spend unprompted.