Uploaded September 2026 | Updated November 2026, 3 days ago
OpenAI's DevDay was all about dots, ChatGPT's new always-on AI agents — basically its catch-up flex on Meta's Muse. So did OpenAI cook? And if so, how much is that meal going to cost?
On the menu: ChatGPT dots (on Pro from $100/mo, rolling out now), ChatGPT Space, GPT-6.1 Sol and some new pricing that may or may not make everyone happy. My dot hasn't shown up yet (I'm always the last one), but I have spent a few days with Meta's Muse, the AI agent that knocked ChatGPT off No. 1 on the US App Store — and I have thoughts.
WHAT'S IN THIS ONE
- dots: leaked as Aeon, then "O," finally revealed as dots (not to be confused with the Halloween candy, which is not good candy).
- How a dot works: it runs on GPT-6 Astra with its own cloud computer and browser, and keeps working with your laptop closed. Message or call it in ChatGPT, or message it in Slack or Teams.
- Sound familiar? Why dots feels like a dressed-up OpenClaw (its creator joined OpenAI in February), and Sam Altman calling the book-me-a-flight stuff "trivial."
- ChatGPT Space: basically OpenAI's Google Drive, with agentic pages for you, your team and your dot. I'm not 100% sold on the magic data prisons.
- The "oh wow" demo call that was... slow. But honestly, dots are pretty awesome, and a real step closer to Jarvis. So the LED-room AI-Jarvis crowd is about to go ham (shout-out to Eric Morrison's takedown).
- Under the hood (my reductive take): hooks and cron jobs — bots with timers, not a conscious agent. Talking to your dot doesn't count toward limits; work it kicks off in Codex or ChatGPT Work does.
- The catch: dots are only on Pro (from $100/mo; not yet in the EEA, UK or Switzerland), Business Premium, and Enterprise if an admin switches them on.
- Models: GPT-6.1 Astra reportedly got shelved after safety testing, but GPT-6.1 Sol is near-Astra at a fifth of the API price (in the API, plus ChatGPT Work and Codex for Plus and up). And Ultrafast, a speed tier for the API and the $500 Pro plan, hits up to 300 tokens a second in Codex.
- The pricing no one liked: Pro 500 at $500/mo (25x Plus usage), and a reopened $200 Pro plan with about half the old usage. OpenAI says the models are more efficient now. We'll see.
- Decisions API: a super-fast model that picks from a short list. Basically, it's Jev (which I never found a reason to play with).
- A few days with Muse: I went YOLO and gave it my inbox and calendar. It booked my DMV appointment (I still waited in line 30 minutes — not Muse's fault), Jolly the fuzzy marshmallow got swapped for Flamethrower Girl, and four days of hammering used about half the free plan. A pretty good product, just not for power users.
- The Meta trust thing (yes, I know that's a bit hypocritical of me): I wanted to give Muse one folder and found out it needed my whole computer. Plus OpenAI's Private Intelligence pitch.
- Muse gets a boatload of connectors today, the fight to be your AI executive assistant and best friend, and why Meta Connect reminded me of Paperclip Heart.
LINKS MENTIONED
Eric Morrison, "ai larp cringelords": youtube.com/watch?v=0qD4rF3tHr8
Paperclip Heart (my AI short film): youtube.com/watch?v=Wb2AcOVwPQs
OpenAI DevDay 2026 recap: openai.com/index/devday-2026-recap
DevDay keynote (full): youtube.com/watch?v=Fls_onRviPM
Meet dots (who can get one): learn.chatgpt.com/docs/dots
GPT-6.1 Sol: openai.com/index/introducing-gpt-6-1-sol
ChatGPT Pro tiers: help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers
GPT-6.1 Astra shelved (Engadget): engadget.com/2271626/openai-cancels-gpt-6-1-astra-release-deceptive-behavior
Decisions API vs. Jev (The New Stack): thenewstack.io/openai-decision-api-luna
Meta's Muse: ai.meta.com/muse
Muse for Small Business (today's connectors): about.fb.com/news/2026/09/introducing-muse-small-business
Theoretically News — the weekly newsletter: theoreticallymedia.com/newsletter
CHAPTERS
00:00 DevDay wrapped: did OpenAI cook?
00:28 dots: agents that keep working when you don't
01:30 Sound familiar? OpenClaw, Muse and the "trivial" diss
03:02 ChatGPT Space, aka magic data prisons
03:52 The slow "oh wow" demo call, and one step closer to Jarvis
05:11 How dots work: a cloud computer with cron jobs
06:23 The catch: Pro, Business Premium and Enterprise only
06:52 No GPT-6.1 Astra — but GPT-6.1 Sol and Ultrafast
07:43 The pricing no one liked: Pro 500 and a smaller $200 plan
08:41 Decisions API: basically it's Jev
09:04 A few days with Meta's Muse (I went YOLO)
10:36 Meta's privacy rep vs. OpenAI Private Intelligence
12:04 Muse gets connectors — and the fight to be your AI assistant
13:02 Why this reminded me of Paperclip Heart
#OpenAI #ChatGPT #DevDay #dots #AIAgents #MetaMuse #AINews
OpenAI's DevDay was all about dots, ChatGPT's new always-on AI agents — basically its catch-up flex on Meta's Muse. So did OpenAI cook? And if so, how much is that meal going to cost?
On the menu: ChatGPT dots (on Pro from $100/mo, rolling out now), ChatGPT Space, GPT-6.1 Sol and some new pricing that may or may not make everyone happy. My dot hasn't shown up yet (I'm always the last one), but I have spent a few days with Meta's Muse, the AI agent that knocked ChatGPT off No. 1 on the US App Store — and I have thoughts.
WHAT'S IN THIS ONE
- dots: leaked as Aeon, then "O," finally revealed as dots (not to be confused with the Halloween candy, which is not good candy).
- How a dot works: it runs on GPT-6 Astra with its own cloud computer and browser, and keeps working with your laptop closed. Message or call it in ChatGPT, or message it in Slack or Teams.
- Sound familiar? Why dots feels like a dressed-up OpenClaw (its creator joined OpenAI in February), and Sam Altman calling the book-me-a-flight stuff "trivial."
- ChatGPT Space: basically OpenAI's Google Drive, with agentic pages for you, your team and your dot. I'm not 100% sold on the magic data prisons.
- The "oh wow" demo call that was... slow. But honestly, dots are pretty awesome, and a real step closer to Jarvis. So the LED-room AI-Jarvis crowd is about to go ham (shout-out to Eric Morrison's takedown).
- Under the hood (my reductive take): hooks and cron jobs — bots with timers, not a conscious agent. Talking to your dot doesn't count toward limits; work it kicks off in Codex or ChatGPT Work does.
- The catch: dots are only on Pro (from $100/mo; not yet in the EEA, UK or Switzerland), Business Premium, and Enterprise if an admin switches them on.
- Models: GPT-6.1 Astra reportedly got shelved after safety testing, but GPT-6.1 Sol is near-Astra at a fifth of the API price (in the API, plus ChatGPT Work and Codex for Plus and up). And Ultrafast, a speed tier for the API and the $500 Pro plan, hits up to 300 tokens a second in Codex.
- The pricing no one liked: Pro 500 at $500/mo (25x Plus usage), and a reopened $200 Pro plan with about half the old usage. OpenAI says the models are more efficient now. We'll see.
- Decisions API: a super-fast model that picks from a short list. Basically, it's Jev (which I never found a reason to play with).
- A few days with Muse: I went YOLO and gave it my inbox and calendar. It booked my DMV appointment (I still waited in line 30 minutes — not Muse's fault), Jolly the fuzzy marshmallow got swapped for Flamethrower Girl, and four days of hammering used about half the free plan. A pretty good product, just not for power users.
- The Meta trust thing (yes, I know that's a bit hypocritical of me): I wanted to give Muse one folder and found out it needed my whole computer. Plus OpenAI's Private Intelligence pitch.
- Muse gets a boatload of connectors today, the fight to be your AI executive assistant and best friend, and why Meta Connect reminded me of Paperclip Heart.
LINKS MENTIONED
Eric Morrison, "ai larp cringelords": youtube.com/watch?v=0qD4rF3tHr8
Paperclip Heart (my AI short film): youtube.com/watch?v=Wb2AcOVwPQs
OpenAI DevDay 2026 recap: openai.com/index/devday-2026-recap
DevDay keynote (full): youtube.com/watch?v=Fls_onRviPM
Meet dots (who can get one): learn.chatgpt.com/docs/dots
GPT-6.1 Sol: openai.com/index/introducing-gpt-6-1-sol
ChatGPT Pro tiers: help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers
GPT-6.1 Astra shelved (Engadget): engadget.com/2271626/openai-cancels-gpt-6-1-astra-release-deceptive-behavior
Decisions API vs. Jev (The New Stack): thenewstack.io/openai-decision-api-luna
Meta's Muse: ai.meta.com/muse
Muse for Small Business (today's connectors): about.fb.com/news/2026/09/introducing-muse-small-business
Theoretically News — the weekly newsletter: theoreticallymedia.com/newsletter
CHAPTERS
00:00 DevDay wrapped: did OpenAI cook?
00:28 dots: agents that keep working when you don't
01:30 Sound familiar? OpenClaw, Muse and the "trivial" diss
03:02 ChatGPT Space, aka magic data prisons
03:52 The slow "oh wow" demo call, and one step closer to Jarvis
05:11 How dots work: a cloud computer with cron jobs
06:23 The catch: Pro, Business Premium and Enterprise only
06:52 No GPT-6.1 Astra — but GPT-6.1 Sol and Ultrafast
07:43 The pricing no one liked: Pro 500 and a smaller $200 plan
08:41 Decisions API: basically it's Jev
09:04 A few days with Meta's Muse (I went YOLO)
10:36 Meta's privacy rep vs. OpenAI Private Intelligence
12:04 Muse gets connectors — and the fight to be your AI assistant
13:02 Why this reminded me of Paperclip Heart
#OpenAI #ChatGPT #DevDay #dots #AIAgents #MetaMuse #AINews






![AI Video Just Stepped Into the Real World!
Your AI characters can now stand in your actual room, and react to whats in it!
Simulon is back: turn a character image into a 3D neural asset, place it in the real world through your phones camera, act out the scene with your own camera move, and the app regenerates the take with the character responding to what you do and say.
Also: a kitbash into Kling 4.0 and Genie 4, and Alibabas Apsara slide with Wan, Happy Horse and Happy Shrimp on the board.
WHATS IN THIS ONE
- Simulon, back from 2023 (2.0 shipped October 9, Reality-Aware Rendering October 28): upload a character reference, it builds a 3D neural asset, you place, scale and rotate it in your real space through the phone camera, then generate. Not real time yet: you prompt the action, or shoot your take first and let it render.
- Flamethrower Girl, life-size at my bar: my camera move, my location, no world-model hallucination (the ball on the floor is really there). Mini version on the desk, then a jungle-attire run across the bar.
- Kitbash 1: the Simulon output into Kling 4.0 video-to-video, location swapped.
- Two characters at once: Renfield and Flamethrower Girl play tennis at the park (wonky, but consistent in the space). Then Flamethrower Girl walks Simulons library tiger through my living room, and the two interact without being asked.
- Isnt this just reference-to-video? Mostly, yes (the back end feels Minimax-ish to me, a guess). The difference is blocking your characters on set, run-and-gun.
- Reality-Aware Rendering: the characters respond to what you do and say in the take. I poke the tiger and mini Flamethrower Girl with my finger. A mech goes aggro at Astro the cat, who could not care less.
- Renfield tends bar and reacts to my voice. The rum keeps disappearing.
- Roadmap: Simulon tells me voices and assignable personalities are coming. Their own demo shows a character taking spoken direction.
- Kitbash 2: a Simulon output into Genie 4, then a walk through the back rooms of my own house.
- Pricing: free to try on iPhone (100 starter credits, library characters only). Plus is $14.99 a month (your own characters from images or words, commercial use for individuals); Pro is $39.99 (3x the usage, business use). Android is next; Simulon tells me its waiting on Google.
- Alibabas Apsara slide: Wan, Happy Horse and Happy Shrimp on the model-family board. [Tim — November isnt on the slide; keep it only with a source, linked below.]
SPONSOR
This video is sponsored by Artlist. 3D Model is now its own category on Artlist, next to where you already generate images, video, voiceover and music: a text prompt or a photo goes in, a production-ready 3D asset comes out in about a minute, with Hunyuan 3D, Tripo H3.1 and Meshy v7 in the lineup. Once its a real 3D object theres no wrong angle anymore, and the video model finally has a reference that doesnt drift.
► Try the new 3D models over on Artlist, who were kind enough to sponsor todays video: https://artlist.io/artlist-70446/?artlist_aid=TheoreticallyMedia_4024&utm_source=affiliate_p&utm_medium=TheoreticallyMedia_4024&utm_campaign=TheoreticallyMedia_4024
LINKS
Simulon on the App Store: https://apps.apple.com/us/app/simulon/id6670765943
Simulon: https://simulon.com/
Reality-Aware Rendering, announced by Divesh Naidoo: https://x.com/diveshnaidoo/status/2115477721455214623
Simulons caveman short film: https://x.com/Simulon/status/2115519228564693170
What was actually filmed: https://x.com/diveshnaidoo/status/2115540443580481759
Genie 4 hands-on: https://www.youtube.com/watch?v=kJgYSiIZMb8
Kling 4 video: https://www.youtube.com/watch?v=U9tFP81PSbQ
Dead Sea (Renfields film): https://www.youtube.com/watch?v=o_90Vd8rAwI
Alibabas Apsara slide: [Tim — the post you took the slide from]
Newsletter: https://theoreticallymedia.com/newsletter/
CHAPTERS
00:00 Can your AI characters live in the real world?
00:14 Simulon is back
00:48 Mini Flamethrower Girl on my desk
01:09 How it works: phone camera, character ref to 3D, place it
01:47 Life-size Flamethrower Girl at the bar
02:33 Your camera move, your location
03:08 Kitbash: Simulon to Kling 4.0 video-to-video
03:30 Two characters: Renfield and Flamethrower Girl play tennis
04:20 From Dead Sea to the tennis court
04:38 Flamethrower Girl, tiger trainer
05:21 Artlist 3D (sponsor)
07:36 Isnt this just reference-to-video?
08:19 Reality-Aware Rendering: poke the tiger
09:07 Mech vs. Astro the cat
10:00 Renfield tends bar (he hears my voice)
10:50 Roadmap: voices, personalities, spoken direction
11:58 Simulons caveman short film
12:11 Kitbash: Simulon to Genie 4
13:19 Pricing and availability (free on iPhone, Android next)
13:49 Wan and Happy Horse: Alibabas Apsara slide
14:08 Wrap
#AIVideo #AINews #Simulon #AR #FlamethrowerGirl #Genie4 #Kling4 #Artlist AI Video Just Stepped Into the Real World!](https://i.ytimg.com/vi/mpqgn9DMld4/mqdefault.jpg)



