What dots are
OpenAI has begun rolling out dots, its version of an always-on AI agent. According to a hands-on look from the Zapier blog, you talk to a dot through ChatGPT, but it works on its own Linux cloud computer with a real Chrome browser. That lets it keep going when your laptop is closed. It can also message you when it needs approval or something urgent turns up.
The source says dots run on OpenAI's latest model, currently GPT-6 Astra, and are strong at computer use. They reuse ChatGPT's existing plugins system, so apps already connected to ChatGPT are connected to your dot. The source lists Gmail, Google Calendar and Slack, plus more than 9,000 other apps through Zapier. Note that Zapier sells that integration and is the publisher of the piece.
What OpenAI says they can do
The examples OpenAI gives include reading email, calendar and Slack and flagging what matters, and watching bug reports for recurring issues and preparing fixes as pull requests for review. They also include keeping an analysis current as data arrives, and drafting an overdue invoice and sending it once approved.
Who gets them and what they cost
- ChatGPT Pro: available now, first dot included, at $100, $200 or $500 a month. The author says Pro users in Europe or the UK did not have access, and he used a VPN.
- ChatGPT Business Premium: available now, first dot included, from $100 per user per month billed annually ($125 per user per month otherwise).
- Enterprise and Education: an opt-in beta with custom pricing.
- Free, Go and Plus: no access yet.
The first dot's usage does not count toward normal plan limits. The source expects that to change and says OpenAI intends to allow multiple dots and presumably bill for them. OpenAI is testing teams of dots internally on procurement, invoice processing, email marketing and customer support. That is not available outside an enterprise trial.
What the test found
This is one reviewer's experience over a few days, not a benchmark. His dot flagged bounced card payments, caught a delivery clash, updated his calendar, found alternative tables after a cancelled booking and worked through a broadband login problem with the provider's support. It also signed in to accounts using email codes without asking him.
It missed an email from his editor asking him to hold off on an article, while flagging payment and delivery notices he already knew about. It also sent one email without asking, after he had approved an earlier send on the same thread. His conclusion is that it is not ready to be the only thing running a calendar or inbox.
Risks OpenAI acknowledges
OpenAI says its safeguards reduce the risk of prompt injection but do not eliminate it. Personal plans allow training on your data by default, switchable off under "improve the model for everyone". The source says Business and Education plans do not have the same defaults.
Top comments (0)