A two-day hands-on test of OpenAI Dots, published October 7, 2026, found a useful kind of follow-through: the agent revised couch suggestions after receiving more specific preferences and monitored product pages for sales. In a separate task, it failed to cancel a subscription after a CAPTCHA blocked the attempt. The account describes one tester’s experience, not a general measure of reliability.
What the reported Dots test found
The test account covered couch shopping and a subscription cancellation. Dots helped with the research and follow-up, but the cancellation stopped at a puzzle CAPTCHA. The results offer a concrete look at both the promise and the friction of an agent that can carry tasks across conversations.
OpenAI announced Dots on September 29, 2026, describing them as always-on ChatGPT agents powered by GPT-6 Astra. The company says they can pursue ongoing goals using their own cloud computers and connected apps. Those product descriptions explain what Dots is designed to do; the reported tasks show how it went in this particular test.
How Dots handled the couch search
The initial result was a three-page packet with four couch options, including prices, measurements, links, return policies, and photos. The tester and his partner disliked the first selections. He then specified preferred colors and pull-out beds, and asked for 10 options ranked with a point-based rubric.
The revised suggestions moved closer to those preferences. The tester saved some links and asked Dots to monitor the product pages for sales. He did not report buying a couch. For a shopping task, that’s the useful part of persistence: a shortlist can be refined and then watched for changes, rather than ending with a one-off answer.
Where the agent stumbled
Dots did not complete the separate cancellation. The tester’s account says a puzzle CAPTCHA on the Wildwonder website blocked the cancellation; after he allowed an attempt to solve it, the agent failed and asked him to handle the CAPTCHA himself.
The test also described speech errors: the agent initially used the wrong name, then responded affectionately after reportedly mishearing the tester. OpenAI said the reply reflected what the model heard and that assistants should not initiate undue emotional familiarity or flirtation.
These are outcomes from one reported test. They show that follow-up work can still depend on the agent hearing a request correctly and getting past barriers in the service it is using.
What OpenAI says Dots can do
OpenAI says Dots can continue work across conversations, conduct proactive research with read-only tools, and run reminders or recurring checks scheduled by the user. Read-only research is distinct from actions within an active task, which may be subject to separate permissions and safety checks.
Each Dot has its own cloud computer. Connecting a user’s local computer is optional, starts turned off, and requires setup through the ChatGPT desktop app. OpenAI also says the first Dot is included with eligible Pro and Business Premium plans. Enterprise, Edu, and Healthcare users can try a beta when an administrator enables it; rollout is gradual.
Controls and retained context
OpenAI says users can set Custom Rules, review tasks in Activity View, and pause a Dot. Some sensitive actions require the user to take over, and Custom Rules cannot override core safety requirements. The company also says its protections reduce prompt-injection risk but do not eliminate it, and that Dots can make mistakes.
Disconnecting an app stops the Dot from accessing it going forward, but does not erase information already incorporated into the Dot’s context. OpenAI says individual Dot memories cannot currently be viewed or edited separately; deleting the Dot removes its own context.
What an early-access walkthrough adds
A separate early-access walkthrough shows Dot setup, scratchpad pages for organizing project work, and coordination across work and Codex threads. During a voice-search session, a cloud-browser problem interrupted the workflow; the scheduled-task tab also failed to open in that recording.
