I Tested OpenAI Dots: The Always-On Cloud Agent Is Slower Than Promised

The agent race is changing shape: models that answer inside a chat window are giving way to permanent helpers that work in the background while you sleep and leave the result on your desk by morning. OpenAI announced this vision on the DevDay stage on 29 September 2026 under the name Dots. According to The Decoder, Dots was introduced as always-on agents that take on tasks day and night on their own cloud computer , carry a browser, and run on GPT-6 Astra. The timing is pointed; arriving only three weeks after Meta's Muse move, the announcement means, according to Wired, a race to carry the proactive-help promise into the mainstream.

The first lesson of setup is the access map: Dots does not open through the same door for everyone. According to TechCrunch, the first Dot is included at no extra charge in eligible Pro and Business Premium plans, and chatting with a dot does not count against ChatGPT usage limits. Against that, according to The Decoder, the European Economic Area, Switzerland and the United Kingdom are currently out of scope for Pro users, while Business Premium is open in every supported region. What happened to the host confirms it: despite paying for Pro on his business account, Dots stayed invisible there while appearing on his personal account, with the workspace side marked unavailable.

Setup and personalization

During setup you choose where the Dot runs: a cloud computer supplied by OpenAI or your own machine. According to the OpenAI Help Center, local computer access is optional and off by default on the Enterprise side, and a dot's cloud machine does not automatically inherit the member's local VPN or browser sessions. The host stays with the cloud option, names his helper Big J, changes its color, and adds a pet icon. Personalization and avatar generation run from the same panel; the host asks for a drawing based on what the agent knows about him, and a fiery orange lobster comes back.

The sleek face of personalization stumbles on the voice-call button. The call key in the top right never answers through the whole test; the host rings twice with no reply and blames the experimental phase. Avatar generation paints a similar picture: the result appears roughly 20 to 30 minutes after the request. The host stresses the response-time gap by noting that a Sonnet 5.5 class model would finish a comparable output in 5 to 6 minutes. The first-impression sentence is blunt: it looks basic and unimpressive.

The research side earns the video's most generous marks. Dots can produce PDFs and lays out an OpenAI Developer Day 2026 summary with its sources step by step. One correction is needed here: the model named GPT-6.1 Soul in the video is actually GPT-6.1 Sol; The Decoder's launch notes spell this economy-band model as Sol. The host also mentions the Astra Ultrafast name; according to the Kingy comparison, that fastest tier is tied to the Pro 500 plan. Everything produced is stored inside the workspace, and the research steps stay visible as you scroll. The Slack connection sits in the same panel.

Computer use and speed

The heart of Dots is a cloud computer that can be watched on a separate screen. When the user asks for the latest news in AI automation, the agent does not move for about a minute and only starts browsing after two minutes. The host calls the gap between this pace and Sonnet-class computer use enormous. Even so, he likes the take-over idea: on a login form or an unclickable button, the user can grab the screen and finish the job by hand. Scheduled tasks, accounts that stay signed in, and document creation inside chat are all pieces of this round-the-clock layout. According to the Kingy comparison, the architectural difference is sharp: while several named Bots on the Grok Bot side share roles, a single primary dot on the Dots side coordinates background agents and delegated work.

The advertisement-versus-reality clash runs through OpenAI's announcement on X: your dot can handle anything from booking a table to your most ambitious work. The host tests the claim by requesting a table for two at an Italian restaurant in Morecambe at 7:30 in the evening. Because his Google account is connected, the agent engages with the booking and, as parallel work, compiles automation news dated 29 September. The table reaches the Vista Italian booking page after a process of about 4 to 5 minutes, and the booking screen opens. Yet the detour through the Terms and Conditions page goes down as pointless slowdown. On the safety frame, according to Wired, approval is requested before sensitive operations, and according to the OpenAI Help Center, users can write custom rules stating which action needs permission, approval, or handoff, while proactive scanning stays limited to read-only tools.

Apps, pricing and verdict

App connections are the messiest chapter of the file. Although 4000 supported apps are mentioned, the interface shows no clear Apps section; the host walks to the extensions screen from settings and recommends connecting through Zapier. According to the OpenAI Help Center, a dot can join supported Slack or Teams workspaces under its own identity and carry the same context across with team members. One more limit is written here: a laptop's browser session does not automatically become a cloud-browser session, and jobs on a personal computer need the device online with the app open. The Kingy comparison records the same boundary as the local-versus-cloud split.

The pricing and region trap is the video's most concrete warning. The host upgrades to a paid plan on his other account for access, then gets told the plan is unavailable in his region. The figures in the Kingy comparison are crisp: the Grok Bot side can be entered from 20 dollars a month through Cursor Pro, while the Dots side starts at 100 dollars a month with Pro 100. According to TechCrunch, deeper work counts against plan quotas with expanded limits for the first month, while tasks delegated to Codex or Work count under their normal allowances. On the business side, Business Premium is priced per seat, and new workspaces require at least two paid seats.

The closing verdict is open and harsh: the host says he will not take Dots and that it is not worth it in this state. In his view, people chasing fast and lean results should look at the Claude side, people wanting a faster and more mature cloud experience at the Grok Bot side, and people wanting open-source flexibility at the Hermes side; he even argues that Mana months ago was better than today's Dots. Fair notes exist too: the visual design quality of website and quiz output gets praised, and the booking is shown completing at the end. Still, the overall picture does not change: slow computer use , a dead call button, and a first image that would not open sit uneasily beside the launch language about peak efficiency.