OpenAI Dots: what an always-on agent changes

On 29 September 2026, at DevDay, OpenAI put an agent inside ChatGPT that keeps working after you close the tab. The capability jump is real. The part that surprises people arrives later, on the bill: a chat model runs when you press enter, and a dot runs when it decides to.

What a dot is

A dot is a named, persistent agent. It holds a goal over time, carries context between conversations, and brings finished work back for review. Five things define it:

• It has its own computer. Each dot runs on an isolated cloud machine with a browser, so the work continues when your laptop is shut.
• It has memory that compounds. A dot is not a fresh session each time. What it learned yesterday is still there today.
• It reaches your tools. More than 4,000 apps connect through the plugin ecosystem.
• It lives where you work. ChatGPT on desktop, web and mobile, plus Slack and Teams.
• It asks before acting. Dots follow ChatGPT's permission model, and you can set rules for when a dot may act alone and when it must stop and ask.

One dot is included with eligible paid plans: Pro and Business Premium, with Enterprise workspaces able to enable it on approval. OpenAI publishes no price for additional dots.

Why it is more capable than a chat window

A chat model is a very good function: text in, text out, one turn at a time. Everything else in a chat product—files, memory, tools—is scaffolding around that function. A dot moves three of those from scaffolding into the product itself.

1. Time. A chat model has no concept of later. A dot does: it can check something tomorrow morning and tell you what changed.
2. A computer. Reasoning about a website is weaker than opening it. A hosted browser lets a dot read the page, click through it and come back with the answer instead of a guess.
3. Accountability. Named agents with their own permissions mean work can be delegated to a specific dot with a specific scope, rather than to one assistant that sees everything.

That is the intelligence story, and it is genuinely new. The cost story follows directly from it.

Where the tokens come from

An agent is not a chat, in four ways

1. It runs without you. Scheduled and proactive work happens on the dot's clock. Nobody presses enter, and nobody sees the cost until the invoice.
2. Memory is input. Persistent context is not free storage. It is re-sent as input tokens on every run, and it grows over the life of the dot. Caching reduces the rate—OpenAI charges $1.00 per million for cached input against $10.00—but it does not remove the volume.
3. Screens are images. Computer use means screenshots. A 1920 × 1080 screenshot is 2,448 input tokens on GPT-6 Astra under OpenAI's published patch formula. Twenty screenshots in one session is 48,960 tokens, before the dot writes a word.
4. Thinking is billed as output. Reasoning tokens are output tokens, at $50.00 per million. A long deliberation costs the same as a long answer.

The arithmetic, in chat turns

A short chat turn is about 1,000 input and 500 output tokens. On GPT-6 Astra that is $0.04.

For agent work there is a published measurement to anchor on. Artificial Analysis, reported by VentureBeat, measured GPT-6 Astra Max at roughly 27,000 output tokens per task on its Intelligence Index—the lowest of the frontier agents it measured, and still 54 times the output of one chat turn. At the published rate, that output alone is $1.35.

The last row is the point of this article. Nothing in it assumes the agent is wasteful. It assumes only that the agent does what it was built to do: run on a schedule. Eight runs of one task is 432 chat turns of output, produced by one agent, on one job, in one day.

This is why a per-seat subscription is the natural way to sell an agent and an awkward way to budget for one. The subscription is fixed; the work behind it is not.

What OpenAI publishes, and what it does not

Published: the GPT-6 Astra API rates used throughout this page ($10.00 input, $1.00 cached input and $50.00 output per million, doubling above 272K input tokens); that one dot is included with eligible paid plans; that proactive background research is read-only, so a dot cannot send messages or change app content on its own.

Not published: any token allowance attached to a dot, the price of additional dots, and how much of a plan's usage a dot consumes. Until those appear, the honest answer to "what does a dot cost per month" is the plan price, and the honest answer to "what does the work cost" is the API arithmetic above.

Price your own agent loop with the token calculator, read the same analysis for Grok Bot and Meta Muse, or compare all three in Dots vs Grok Bot vs Muse.