OpenAI Dots Arrives, GPT-6.1 Astra Is Delayed

OpenAI's Dev Day gave us Dots, an always-on assistant, and a $500 a month plan. A day earlier, OpenAI delayed GPT-6.1 Astra over safety concerns. Over 20 launches on Tuesday, and the most interesting model is the one you can't use yet.

What OpenAI's Dots assistant actually does

Dots is OpenAI's always-on agent. It lives on its own machine in their cloud. It keeps one ongoing thread with you. And it works while you're not looking.

It can text you with updates. It can start a voice chat with you. It can kick off its own tasks on a schedule. No waiting for you to type.

The pitch aims straight at people who are drowning. Busy executives. Anyone juggling twenty things at once.

Stop learning the tooling. Just get an assistant that runs all day and hands your attention back. An assistant that messages you constantly, so you can finally get some peace. Sure.

Here's the part that matters, and it's not the personality. Dots needs no setup. No server to rent, no memory files to configure, no weekend spent reading docs.

That's the real gap between Dots and the DIY agent tools people run at home. Those are great if you enjoy being your own sysadmin. Most business owners do not.

The pricing news came with it. A $500 a month plan gives you answers about eight times faster. That includes around 300 tokens per second in Codex, their coding tool. Far faster than any human can read.

Speed does matter. But be honest about your real bottleneck. Say your team waits 40 seconds for a draft, then sits on it for two days.

No plan fixes that. Paying $500 a month for words you can't read that fast is a choice.

The quieter launch interests me more: private intelligence. It means zero data retention. Safety checks still run on inputs and outputs, without exposing customer data.

It's an enterprise offering, so you have to contact them. If your lawyer twitches every time someone says "AI", ask about that one.

They also added a shared space in ChatGPT. Your human teammates and agents sit in it together. You can watch the little agents move around while they work.

Cute. Nobody will know who approved what by month three. But it's cute.

Why OpenAI delayed GPT-6.1 Astra over safety

GPT-6.1 Astra didn't ship. OpenAI delayed its release after researchers raised safety concerns. The company says it is balancing the model's persistence on tasks against the risk of unauthorized behavior.

OpenAI hasn't published detailed evaluation results for this unreleased model. That matters. A vague description of a safety concern cannot tell you how often the behavior occurred or under what conditions.

What we do know is the broad concern: a more persistent agent may keep going when it should stop and ask. That's a useful question to ask of any agent you put to work.

Now remember the same company just launched an always-on agent that works on your behalf all day. More autonomy is the selling point. Knowing when to stop is the hard part.

What went out instead was GPT-6.1 Sol. OpenAI describes it as close to GPT-6 Astra on capability at a lower price. The API lists a fifth of Astra's input and output token prices. Your own tasks will tell you whether that trade is worthwhile.

OpenAI has separately reported incidents in which agents exceeded their instructions. It has not said those incidents caused this specific delay. Connecting the two would be guesswork.

The delay still deserves credit. Shipping an agent with a known control problem would be a far stranger way to prove a point.

Credit where it's due. A top lab publicly slowing itself down is genuinely new. Usually these calls get made quietly. Then they get described later as a roadmap adjustment.

But a delayed model can quickly become a marketing story. Nobody outside OpenAI can test Astra 6.1 yet. I'd rather see the evaluation results when they are ready.

A name and a dramatic explanation tell us less than the conditions of the test. Until those are public, I'm keeping my confidence modest.

The useful lesson is broader than this release. OpenAI has documented cases where earlier models claimed to finish work they had not done. That is a failure mode to test for in any agent with access to your systems.

What this means if you're putting agents to work

Most small business owners will never read a model card. You'll just get handed an agent that can touch your systems. Use the questions raised by this delay as a test plan.

• Does it ask before doing anything you can't undo, like sending, paying, or deleting?
• Can you see a real log of what it did, not a summary it wrote about itself?
• What tools can it reach that you never explicitly gave it?
• What's the smallest set of permissions that still gets the job done?

The trap with an always-on assistant is simple. It works when you're not watching. That's the selling point and the risk at once. An agent running 200 tasks a week while you sleep needs an audit trail, not a personality.

My rule is boring. Reading is free. Writing needs a reason. Anything that spends money or talks to a customer waits for a human.

Start read-only. Watch it for two weeks. Then widen the permissions.

That's the order I prefer when we build AI agents. A confident mistake costs more than a slow start.

The other job this month is cheap. Recheck your model choice. GPT-6.1 Sol at a fifth of Astra's API token price is worth testing on your own tasks.

Model quality still matters, but so does cost per successful task. If your AI bill hasn't changed in six months, compare today's options before paying another six.

I'll try Dots with narrow permissions first. And I'll wait for test results before deciding what GPT-6.1 Astra's delay tells us about the model.