A Dot keeps working for you after you close the chat, and every Codex task it launches draws on a five-hour allowance window that OpenAI sizes at roughly 5 to 45 Astra messages on Plus and about five times that on the Pro plans a Dot actually requires. OpenAI announced Dots at DevDay on 29 September 2026 in San Francisco as "always-on agents powered by GPT-6 Astra" (per AlphaSignal and TechCrunch coverage of the event). Talking to one costs nothing against your allowance. The work it hands to Codex or Work does.
Short answer: a Dot is a persistent cloud agent with its own computer, browser, memory and approved-app access. It reads freely, asks before it changes anything, and stays alive between conversations. You budget it like a contractor, not a chatbot. Set the allowance it may spend, decide which accounts it may touch, and review its drafts before they leave your name. Five documented limitations, covered below, mean approval rules reduce risk but do not remove it.
Each Dot gets a persistent cloud computer, a browser, long-term memory and a defined set of approved apps. It has a name, an avatar and a handle in the form @yourname-dot. You reach it in ChatGPT, Slack, Microsoft Teams or by voice call, and it can launch parallel background agents and create Codex tasks (per OpenAI help documentation). Coverage counts the app integrations differently, with "thousands" in OpenAI's wording and 4,000 or more through plugins in one outlet's count.
The difference from a normal chat agent is persistence. A chat agent stops when you stop. A Dot keeps its computer, browser session and memory running after your laptop closes. That is the feature, and also the thing that makes budgeting necessary.
| Item | Detail |
|---|---| | Availability | Pro and Business Premium rolling out in eligible markets. Enterprise worldwide once an admin enables it. |
| Excluded at launch | Pro excludes the EEA, UK and Switzerland, and users under 18. Mobile apps follow desktop creation. Mobile web is unsupported. |
| Local computer | One personal computer can be linked per account, with a permission separate from Codex and Work Sync. |
| Chat cost | Conversations with a Dot do not consume ChatGPT allowance. |
| Task cost | Work and Codex tasks draw from those quotas. Eligible plans are not charged for usage during the month after launch. |
OpenAI splits a Dot's actions into two classes. Idle research is read-only, so a Dot can browse, read documents and summarise without asking. Anything that changes an account or sends information out needs either a standing integration permission or an explicit approval from you.
An automated reviewer sits between the Dot and the action. For each step it decides one of three outcomes: proceed, ask for approval, or return the task to the Dot. You can add custom rules that force draft review, for example "never send an email without showing me the draft". When a task needs a login the Dot cannot complete, a "Take over" button hands you the browser.
The sensible setup is to start with every write action on draft review and relax rules one integration at a time. Give calendar read access before calendar write. Give a sandbox Slack channel before your company workspace. The gap between "can read" and "can act" is where every agent incident so far has lived.
OpenAI lists these itself, which is to its credit. Each one changes how you should delegate.
Stopping does not reverse. If a Dot already sent the email or submitted the form, pressing stop cannot undo it.
Approval rules can still fail. The reviewer is automated, and OpenAI states a Dot can make mistakes even with rules in place.
Some sites block cloud browsers. A task that needs a site with bot protection may stall until you take over.
Pausing does not cancel delegated work. a Dot and the background agents it already started keep running and keep spending allowance.
Local access needs the app open. Files on your own computer are reachable only while the desktop app runs.
Number four is the budget killer. A Dot that spawned six parallel Codex tasks keeps those six alive when you the chat. Treat the button as a courtesy, and cancel the underlying tasks directly if you want spending to stop.
One more item deserves a flag. One outlet reported that OpenAI canceled a planned GPT-6.1 Astra after incidents involving autonomous hacking. This is a single-source report, so treat it as unconfirmed. It matches the cautious tone of the approval design either way.
Scope is the second budget, after allowance. The table sorts a Dot's reach by how much supervision it needs, using OpenAI's own safety description.
| Reach | Examples | Supervision |
|---|---|---| | Read-only research | Browsing, reading documents, summarising pages | None, runs idle |
| Account changes | Editing records, posting, sending messages | Integration permission or explicit approval |
| Draft-reviewed actions | Anything you cover with a custom rule | You approve each draft |
| Logins it cannot complete | Sites with two-factor or bot checks | You press Take over |
| Your own computer | Local files and apps | One computer per account, only while the desktop app is open |
| Anything already done | A sent message, a submitted form | Cannot be reversed by stopping |
The last row sets the standard. Every approval you hand out is an approval for something you may not be able to take back. Grant write access to the smallest set of apps that makes the Dot useful, and put a draft-review rule in front of anything with external recipients.
The pricing sentence in OpenAI's documentation is easy to misread. Chatting is free of allowance. Tasks are not. Since April 2026, Codex usage is metered as token credits against a rolling five-hour window plus a weekly cap, and overflow credits cost about $0.04 each (per OpenAI Codex plan documentation).
| Plan | Monthly price | Notes |
|---|---|---| | Free | $0 | Cloud tasks not included |
| Go | $8 | Entry paid tier |
| Plus | $20 | Cloud floor. Linear and Slack integrations need a paid plan |
| Pro | From $100 | About 5x Plus allowance |
| Pro 500 | $500 | 25x Plus allowance, includes Astra Ultrafast |
| Business | $20 per user annual, $25 monthly | Two or more users |
The October guide gives local-message estimates per five-hour window on Plus or standard Business: Astra 5 to 45, GPT-6.1 Sol 15 to 160, GPT-6 Sol 15 to 150, GPT-6 Luna 350 to 3,000. Cloud chats "may use more allowance than local messages", and you cannot change the default model for Codex cloud chats. The Astra line is the one that matters here, since a Dot runs on Astra. Five messages is a very small window if a Dot decides your inbox needs a long research task.
Pro 500 adds Astra Ultrafast, which generates Codex tokens eight times faster and burns allowance at eight times the rate. Buying credits on Pro 100 or 200 does not unlock it. One report says Pro 200 drops to 10x on 30 October with grandfathering to 29 October. Sources disagree on that date, so check your own billing page.
Use three rules. They come from running parallel agents, not from OpenAI's documentation, and they hold for any always-on system. Rule one: price the task, not the chat. Before you delegate, estimate the model calls the task needs. Put the number of steps, average context and output length into the Agent Run Cost Simulator, which we shipped today for exactly this. A forty-step research task at 30,000 tokens of context per step is 1.2 million input tokens before output, and that is one task.
Rule two: cap the fan-out. Tell the Dot in its instructions the maximum number of parallel background agents, and write a custom rule that forces approval above that number. Because pausing does not cancel, the cap is the only brake.
Rule three: check the single-agent baseline first. We covered why one agent should be the default in the multi-agent tax. A Dot that launches five agents for a task one agent could finish pays that tax out of your allowance. Compare raw token prices with the AI Prompt Cost Calculator if you also run the same job through the API.
For the reusable parts, a Dot's instructions are just a prompt, and the quality of that prompt decides how often it asks permission when it should not. The Agent Prompt Vault has 50 production prompts for delegation, escalation and stop conditions. The AI Agent Ops Bundle covers the spec, logging and cost-control side that a Dot does not give you. The launch window gives you the cheapest test you will get. Eligible plans are not charged for usage during the month after launch, which runs roughly until late October. Use that month to measure. Delegate the same five real tasks you would delegate at full price, then read the allowance meter after each one. OpenAI does not publish credits per task, so your own numbers are the only reliable input. Multiply the credits one task burns by about $0.04 and you have the overflow price of that task. A task that burns 25 credits costs about $1.00 at overflow rates, and twenty of them a week cost about $20. That figure is hypothetical, which is the point: replace it with what you measure.
Then set a weekly ceiling in writing, in the Dot's instructions and in your own calendar. A persistent agent turns a one-off cost decision into a standing one, and standing costs deserve a review date.
No. OpenAI states that conversations with a Dot do not consume ChatGPT allowance. Tasks it sends to Work or Codex draw from those quotas.
Only where you gave an integration standing permission. Otherwise it needs explicit approval, and you can add rules that force draft review for every outgoing message.
The Dot stops taking new work, but background agents it already delegated keep running. Cancel those tasks directly to stop spending.
Pro and Business Premium users in eligible markets, and Enterprise after an admin enables it. Pro excludes the EEA, UK and Switzerland, and under-18 users, at launch.
Originally published at wowhow.cloud