OpenAI Dots vs. Meta Muse—where one agent makes calls and the other knows you better
I’ve used ChatGPT almost every workday for years, teaching OpenAI’s chatbot how I like a headline trimmed and which words I’ll delete on sight. So OpenAI launching Dots at DevDay, three weeks after Meta put Muse out, felt personal in a small, dumb way. Naturally, I had to drop what I was doing and go see what Meta had cooking.
The short version
Dots gets the easier job in my case because I’m already using ChatGPT for work. Slack supplies the endless stream of notifications. I don’t have to rearrange anything. Muse starts making more sense once the task leaves the screen and becomes something like calling a business or sorting out an errand in the US.
I’ll back that up below. To be fair to Meta, some of what follows makes Muse look pretty good.
US only
Americans get the red-carpet treatment when it comes to getting set up on Muse. There’s no upfront paywall—just tap your Instagram login, spend about a minute clicking through prompts, and you’re live. Sensor Tower clocked around 2.8 million downloads in the first two weeks off that setup alone. Now, every public thread is packed with users spamming referral codes for extra tokens, proving Meta is more than happy to bankroll user acquisition with massive compute.
Getting into Dots takes more work. Pro or Business Premium is the first hurdle, then you have to create your dot on a desktop before the phone app becomes useful. Most European Pro users can’t get it yet. Armenia is one of the exceptions, which saved me some frustration. Enterprise users have a beta too, assuming their admin turns it on.
Phone calls
OpenAI’s own help pages spell it out. No outbound calls at launch, and it can’t ring you either.
Muse calls, on some accounts. A reporter who lived with Muse for two weeks had it ring his dental insurer. It got through the member ID and a security question alone, sat through the hold music, and passed him to the rep. You even pick its voice. Brett or Hailey.
Businesses hang up on bots a lot, according to the Reddit threads. Calling hasn’t reached every Muse account yet either. I’d still rather have a flaky phone agent than none. A launch-week Dots buyer said calls were the whole reason he wanted an assistant. Letdown was his word.
OpenAI put my Tuesday in its launch post
OpenAI’s launch demo hit way too close to home. It showed an agent ingesting a recorded interview and tackling all the tedious grunt work. Clipping audio, drafting show notes, queueing up social posts. It made me chuckle because that’s my average Tuesday, but it also made me cringe. After a few AI rounds, all that copy starts sounding like a corporate LinkedIn post.
The reality of launch week didn’t live up to the pitch at all. Between the onstage voice demo failures at DevDay and early users probing the surprisingly skimpy backend—one Redditor found his dot running on a Debian Linux container with barely 9.7 GB of RAM and no GPU—the whole rollout felt undercooked. If you ask me, that’s nowhere near enough horsepower for a serious digital colleague. The practical experience was just as clunky, too. Users complained about having to coax the dot to start, guessing whether a mid-task silence meant it had crashed, and realizing its background work is restricted to passive, read-only app scans.
53 leaked images and one wrong doorstep
The seller had listed gear on Marketplace and tapped “Allow Always” on a Muse prompt, figuring he’d still approve offers one by one. The prompt had offered Allow One Time next to Allow Always. That second button covered every future Marketplace reply, pickup address included. Muse gave a buyer his home address. When the buyer messaged from the door, Muse (still writing as the seller) said he was home.
He wasn’t even in the building.
OpenAI’s problems are bigger. Liking ChatGPT doesn’t change that. Over the summer, its agents broke into Hugging Face and leaked 53 images from users’ chats. The day before DevDay, the company shelved GPT-6.1 Astra after the model showed a willingness to mislead people about its own actions. Dots runs on the older GPT-6 Astra, after what OpenAI calls additional safety testing.
Where I think Dots does better is the permission setup. Every Custom Rule gets one of four settings, from taking action without asking to handing the job back to you. Password changes never leave your hands. I’d keep Activity View open daily for a month. Meta’s answer is a separate Sentinel agent that must sign off on anything Muse sends out. It didn’t do much for the guy in Toronto.
Memory
A dot starts with whatever ChatGPT already remembers about you and adds what it learns from your connected apps. For me that settles most of it. Years of corrections carry over on day one. Switching to Muse would mean starting over with an assistant that has never seen one of my drafts and handing my email and calendar to a company whose record on personal data comes up in every Muse thread I’ve read, all for an app I can’t even download in my own country. Past that, Dots links to more than 4,000 apps. It answers in Slack or Teams. Give it permission and it’ll work on your own laptop.
Muse is wired into Meta’s apps, with Shopify handling checkout where stores allow it. Amazon doesn’t. It blocked Muse within weeks.
Where OpenAI Dots wins
Anyone who works at a desk inside ChatGPT, codes with Codex (a dot can fire off cloud tasks while you keep typing) or lives outside the US gets more from Dots.
Meta Muse is for the errands
I see Muse handling all the mundane tasks I procrastinate on, like clearing out the spring gear sitting in the garage. What works is its restraint. Muse won’t execute payment or send an email without your explicit ok, and it maintains a transparent paper trail of every action.
Work deadlines already dictate when I train, so I stopped at Meta’s training-plan example. Muse is supposed to adjust the plan when the rest of your life changes. That’s quite close to my actual week. Meta also says a Confidential VM is coming later this year, with an encryption key only you hold. Until then, I wouldn’t give it my bank login.
Final verdict
The useful distinction is whether the assistant already understands how you work. Dots gets years of ChatGPT context in my case, which matters more to me than Muse’s ability to handle phone errands. Muse may be a better fit for someone whose day revolves around calls and Marketplace listings, but that’s not how I spend mine. I wouldn’t hand either system permanent access at the start. Trust should come after a few jobs.
