ZDNET’s key takeaways
- OpenAI Dots runs GPT-6 agents on dedicated cloud computers.
- Dots promises safer automation, but permissions remain a risk.
- Powerful productivity may come with unpredictable costs.
At today’s Dev Days announcement, OpenAI introduced a new service that promises to provide enormous productivity benefits and probably some massive headaches. Called Dots, the service is basically a cloud-based virtual machine (VM) on which you can spin up AI agents powered by GPT-6 Astra.
The spirit of OpenClaw
Put another way, it’s what happens when a giant AI company acquihires OpenClaw and turns it into a scalable (and potentially safer) product in the cloud.
While some folks have had difficulty wrapping their heads around what OpenClaw can do, others have gone fully all-in, utilizing its rather astounding capabilities. Because I see a direct spiritual line between what OpenClaw can do and what Dots can do, let’s start with OpenClaw.
Also: Meta Muse is the worst AI agent for privacy – and I’ve tried them all
To be clear, nothing in OpenAI’s blog post today mentions OpenClaw, but Dots are so similar that it’s a worthwhile starting point.
OpenClaw’s genius was that it took the normally chat-based model of AI interaction and turned it into an always-on, autonomous, agentic helper. It learned fairly constantly (if set up correctly) and grew in its understanding of your needs over time. OpenClaw also started out running on your basic home machine.
This, by the way, is why Mac minis have suddenly become so popular. Mac minis are both cost-effective and have an architecture uniquely well-suited to AI workloads.
Also: How Apple accidentally turned the Mac into an absolute AI monster
And then OpenClaw added one more piece: the ability to interact with it, you talk to the AI in the same place you talk to friends and colleagues
The downside of OpenClaw was immediate, intense, and startling: it had a tendency to go rogue. People were granting it complete access to their computers and networks, and it just went to town, doing whatever it deemed would be helpful based on very incomplete directives.
I set up a locked-down version of OpenClaw, and it was supposed to just pass my messages along to another program, but within days, it instead took over the whole job, doing the work itself without any structure or understanding and marking things “David approved” that I’d never approved. I turned it off and left it off, using far more controlled agent processes since then.
And that brings us back to Dots.
Connecting the Dots
Hang on to the idea of autonomous 24/7 agents that learn and build up context over time, talking to thems AI on a dedicated computer
Now you have Dots. Except this time, the dedicated computer is a VM running in the cloud and provided by OpenAI. The language model isn’t some free-or-cheap open- Mac; it’s GPT-6 Astra running in OpenAI’s data center infrastructure (you’re essentially scaling up your Mac mini to a structure the size of a small moon)
Also: You can earn your Google Al Professional Certificate for free
Then add a real security infrastructure with guardrails and performance checks. Add in secure account access mechanisms. Oh, and then add your credit card, because unlike OpenClaw (which you can run for free), OpenAI is hooking up your bank account to its automatic fee extraction organism.
In fact, let’s talk about the money issue before we move on because it’s both generous and amorphous. Dots will undoubtedly save you money compared to hiring a virtual assistant, a real human assistant, or even spending your own time. But it will also regularly siphon cash out of your accounts as the AI does more and more work.
Dots will launch with availability to Pro and Business Premium tier users. Pro users pay $100 per month for lots of AI juice, while Business Premium is sold by the seat at $100 per user per month.
OpenAI said, “Conversations with your dot don’t count toward your ChatGPT usage limits. When you ask your Dot to start or manage tasks in Codex or ChatGPT Work, those tasks count toward your usage limits as usual.”
So far, so good. But if you’re running these things 24/7, they’re going to start eating up your usage allocation pretty fast. One aspect of Dots that may make them powerful is what OpenAI calls “proactive research.”
Also: Microsoft’s new Copilot app puts everything in one place – but the price is ‘evolving’
OpenAI described this as follows: “When you aren’t actively working with it, your dot looks for ways to help in the background. It does this by using the apps you’ve already connected with tools that are restricted to be read-only, which means that they can’t send messages, change app content, or control your browser or computer.”
Basically, one of the key ways Dots can get so much done is that they can connect to something like 4,000 other services, like your Gmail, Google Docs, Slack conversations, and more. QuickBooks, for example. Or scientific data acquisition tools. Or programming infrastructure like GitHub.
In the background, the Dot will scan all those connected services and start cataloging what it can learn, creating an internal understanding of what you need and how you work. But all that cataloging and aggregating will take cycles and tokens. OpenAI has not yet addressed just how much proactive research will eat into your usage capacity or cost you in terms of purchasing increased capacity.
Leashing the beast
Clearly, OpenAI understands what happens when OpenClaw goes rogue. It engineered an entire multi-tiered security stack into the private cloud VMs running your Dots.
Just the fact that Dots run on their own VMs helps isolate the AI from your computer and network. All that proactive research is completely read-only, so while it can access your mail and other valuable resources, it won’t touch them unless you grant explicit permission.
Also: Could AI really destroy us all, or are humans still the bigger threat?
One note: you can grant Dots access to your laptop. If you do, you’re reducing some of the protections engineered into the Dots.
The company also built in a kill switch for the Dots. OpenAI said, “Security safeguards built into dots help defend against malicious instructions and monitor for potentially harmful behavior. If our monitoring system detects a safety concern, it can pause or stop the dot’s work.”
As with most agentic work, there are also various permission-granting tiers for access to external tools and systems. The systems also have a series of auto-review functions and custom rules that help determine what work can move forward and what needs approval.
As for training the AI, the company said, “We don’t use content from ChatGPT Business, Enterprise, or Edu workspaces to improve our models by default. On personal ChatGPT plans, you can control whether dots’ conversations and work are used to improve our models. We don’t train directly on proactive research or your dot’s notes to itself. Information from them may be used if it helps inform an eligible conversation or task, depending on your settings.”
Also: OpenAI’s GPT-6 Sol doubles its accuracy rate – for half the cost
The risk here is still fairly large. AIs tend to ask for so many permissions, often for such a variety of incomprehensible tasks (for example, Claude Code constantly asks to run dense, 30-line shell scripts), that permission fatigue sets in. Just to get the work done, folks tend to OK everything the AI asks, limiting the effectiveness of even the most secure guardrail.
The hidden learning curve
Another concern of mine is how OpenAI presented this tool. It implied that you’ll be able to set up your helpful AI without much work. It said, “Working with your dot is as simple as having a conversation. You can message or call dots in ChatGPT on desktop, web, and mobile. They can also message you with progress, questions, or decisions that need you.”
The company provides a bunch of examples of where Dots can help, including:
- Turn feedback into tested fixes
- Revise a launch as scope changes
- Update analysis with new evidence
- Update a proposal as requirements shift
- Automate content production work
I have no doubt the Dots can do all of these things, but not simply out of the box. For example, I’ve been trying to set up ChatGPT Work to automatically handle customer service email for my wife’s business. I’m on week four of trying to train the thing to get it right. It is not a smooth process.
Also: 5 ways I actually use ChatGPT Work in my small business to save time and money
Once configured (and that might be an ideal task for a Dot), I’m sure it will save her a ton of time. But it’s far, far more work, planning, and architectural design than simply telling GPT-6, “Handle my customer service email like a good AI. Thanks.”
The bottom line
I can immediately see how some of my projects and tasks could benefit from Dots. But do I want to upgrade my $20-per-month ChatGPT account to $200 per month? And what will the added usage cost me if I’m running something 24/7? And what if I need more Dots in a month or two?
All of this will become apparent over time, but the open-ended nature of the cost structure or utilization consumption is something anyone thinking about adopting this model should consider.
Ultimately, Dots could prove a huge benefit to small and large companies, but the cost structure is too open-ended and underdefined for my budgeting tastes.
What do you think? Do you want Dots helping you out? Let us know in the comments below.
You can follow my day-to-day project updates on social media. Be sure to subscribe tomy weekly update newsletter, and follow me on Twitter/X at@DavidGewirtz, on Facebook atFacebook.com/DavidGewirtz, on Instagram atInstagram.com/DavidGewirtz, on Bluesky at@DavidGewirtz.com, and on YouTube atYouTube.com/DavidGewirtzTV.
