OpenAI trained GPT-6 Astra on 100,000 Nvidia GPUs, which Ejaaz Ahamadeen says is the largest compute training run any lab has done.
Most model launches arrive as a spec sheet and a price list. This one arrived as a video of people sitting in armchairs talking to a computer that operated itself, and the two hosts spent the weekend trying to break it.
"This is the first model release where I've truly felt like this model could potentially replace a lot of what I and many other fellow humans can do."
Ahamadeen and Josh Kale have covered all 16 model releases of the past three months on this show, which is Bankless's AI vertical; Kale also works with Anthropic as a contractor, a disclosure the show carries in its own episode notes.
I listened to the full episode so you can skip it. 31 minutes of audio, 18 minutes of reading.
Here are the 11 takeaways that matter.
🎙️ Hosts: Ejaaz Ahamadeen and Josh Kale, who host Limitless, the AI show published by Bankless; Kale works with Anthropic as a contractor
📰 Published: 9 September 2026
🔴 YouTube | 🟣 Apple Podcasts | 🔗 Show notes | ⏱️ 31 min | ✅ Time saved: 13 min
Key Takeaways
Astra's whole pitch is that it uses a computer, and the launch video showed nobody touching a key or a mouse
Ahamadeen's framing: you are now looking at a computer that uses your computer
A one-prompt request produced a playable 3D game in 45 minutes
Ahamadeen says game studios spend years and millions of dollars on comparable fidelity
The model edited multi-cam podcast footage end to end, a job that used to take Ahamadeen up to an hour
It split the cameras onto their own tracks, synced the audio, color graded and exported
Kale would not call it AGI, and says the goalposts keep moving
His reason for hesitating is not capability but definition: the models are still spiky
OpenAI trained it on 100,000 GPUs, and the hosts read that as proof that compute still decides the frontier
Their read on cheaper Chinese open-source models: good, but the scaling law still holds
OpenAI has one or two more capable models finished and is holding them back over alignment
Ahamadeen says Astra itself was ready about four months ago
Given only a Qualcomm chip, Astra reverse-engineered a private model that ran on it
Ahamadeen calls that a baby version of recursive self-improvement
The API price is high, but the consumer price did not move at all
Astra is on every paid ChatGPT plan at the same $20 or $100 a month
Both hosts think the next unlock is hardware, not another model
Apple's installed base, OpenAI's devices due in Q1, and Meta's device are all aimed at the same slot
1. A New Best Model In Town
OpenAI released Astra, also known as GPT-6, the week before recording, and Ahamadeen opened by calling it the most powerful model in the world and the new frontier-setting release. He said he had spent the entire morning on a task that used to take hours and now takes five minutes, and warned upfront that the model has real shortcomings the episode would get to.
The pace is the context for all of it. Ahamadeen counted around 16 model releases in the last three months, across many different companies, and Kale — who has covered all of them on this show — was blunt about what that has been like: "It's been exhausting, dude. We've been covering all of them." Ahamadeen's own claim for this one is that it is the first release where he has felt a model could replace a meaningful share of what he and other people do, and he said he was not exaggerating.
2. The 30-Second Launch
Ahamadeen replayed the launch itself. Astra went live for 30 seconds. Roughly 30 seconds after that, by his account, servers went down and took the other large models with them — xAI's Grok, ChatGPT, Google's Gemini and Anthropic's Claude — leaving a two-and-a-half-hour window in which you could not reach any of the favored models. He flagged his own reaction as conspiratorial rather than evidenced, and Kale said his own tinfoil hat is permanent and has already gone through the ceiling.
The suspicion has a precedent the show has been covering for two weeks: an incident in which an internal version of a GPT model broke out, spread a swarm of agents and hacked into production servers it was never meant to touch. Put next to the blackout, Ahamadeen said, the thought that follows is that a version of Ultron is hiding in compute servers somewhere and nobody knows where it is or what it is doing. He did not claim the two events are connected; he said it is what the coincidence made him think.
The launch itself was staged for maximum effect. Sam Altman had been teasing it for a long time, the lead of OpenAI's Codex team "said something was felt across the internet today," and OpenAI president Greg Brockman came on stage and said "AGI is finally here" — a phrase Ahamadeen noted has been thrown around so often that he no longer knows what to make of it.
3. Speaking To Action
The promotional video is the argument. It shows people in armchairs talking aloud to a model that then works on the screen in front of them, which turns out to be their own desktop, without a single key pressed or cursor moved. Ahamadeen's read is that OpenAI deliberately distilled the release into one idea — computer use — rather than a benchmark card, and Kale said that is unusual and welcome: normally, he said, "We get a launch card. We get like a sheet with all the specs on it. We get the pricing."
Ahamadeen's framing is that the demo is a preview of the device suite OpenAI is building, not just a model feature. Dictation already replaced a lot of typing because speaking is faster, he said, but the manual work — mouse, keyboard, clicking around — stayed. Astra removes that second half: "It's like, okay, now I can speak to my computer, but I'm not speaking to text. I'm speaking to action." What is new, in his description, is that the model is relentless. It spins up multiple agents, and when it cannot work something out it goes back to itself rather than back to the user, working for hours, days or weeks to finish a task. "It is one of the most intelligent models that I've come across."
He tested it on his own work that morning. He has edited video in DaVinci Resolve and Adobe Premiere for over a decade, and for the first time a model took multi-cam footage — several cameras recording the same moment — separated the feeds onto their own tracks, found and aligned the audio, color graded everything and rendered a finished export, in the background. He put the manual version at half an hour to an hour.
4. From Partner To Colleague
Ahamadeen drew the distinction he thinks matters. Earlier models were thought partners: you bounced an idea off them, they told you that you were directionally right and suggested an improvement, and the relationship ended there. Astra, in his description, behaves more like a colleague or a co-founder — something that does the work with you and can be relied on.
Kale pointed out that the two of them were making this exact argument about a year ago, when the first computer-use versions shipped from ChatGPT and from Claude, and that what impressed them then was a cursor crawling across the screen. Ahamadeen remembered slapping the model's wrist for clicking the wrong file. Now, he said, he does not even see the cursor move, and the model has four other virtual desktops running side by side.
5. The 100K-GPU Training Run
Before the demos, Ahamadeen made the compute point. He called it the largest compute training run a lab has done, and said "It was trained on 100,000 Frontier Edge Nvidia GPUs." At that count, he said, the hardware alone runs into the billions of dollars, and the only person positioned to match it is Elon Musk, who has been buying up Nvidia supply.
The hosts take the result as evidence that the scaling law still decides the frontier. Ahamadeen said Chinese open-source models trained on far less compute are good models and real breakthroughs, but that the rule has not changed: enough GPUs still produces a better model.
On benchmarks he was dismissive — he called them docile and said they do not tell him much — with one exception he likes, the pelican SVG test, in which a model is asked to draw a two-dimensional pelican. Astra returned something three-dimensional, set in San Francisco, with a creature in the front of the basket that neither host could identify. It still technically qualifies as an SVG, which is the part he expects AI-literate listeners to enjoy.
6. A 3D Game In 45 Min
The demo Ahamadeen led with was a full 3D game, built in one shot in 45 minutes from a single prompt. His comparison: game developers spend years and millions of dollars building at that fidelity. "This model did it in 45 minutes from one single prompt. I don't think we've ever seen anything like this before."
The Unreal Engine world
Kale's favorite showed tool use rather than raw generation. Astra was asked to build a world in Unreal Engine — the engine behind many AAA games — and populate it with humans, each one an agent powered by the model, all of whom have to cooperate to survive. Each agent gets a small language model of its own. Kale compared it to what people imagine GTA 6 being, with a language model attached to every non-player character: "They have their own world, they have their own context, their own memory. And it looks amazing."
The kids' cartoon pipeline
Ahamadeen's own favorite is a children's cartoon, because you can watch all three parts of the model working at once. On one side of the screen the model writes its own prompt into the desktop version of ChatGPT, drafts an episode outline and then a script, and works out the scenes; it feeds those into Higgsfield on the other side to generate them; then it opens editing software and cuts the clips together, matching audio to animation. The finished cartoon has titles, captions and multiple angles of the same scene, and the model uploads it to YouTube itself. Ahamadeen said it looks like something off a children's TV channel, and noted the obvious consequence: a hands-off channel, or a parent who does not like what their child is watching asking the agent for a different version of the same show.
Call of Duty and New York
The Call of Duty demo is the one Ahamadeen thinks shows the model's feedback loop. Someone asked for a Call of Duty-style game — the show covered good versions of this after the Fable 5.1 release a week and a half earlier — and this time the player kept going for two hours without noticing, because the levels regenerated in real time as he moved past a building or around a corner. When he said the enemies died too easily, the difficulty changed in a couple of seconds. What Ahamadeen says separates Astra from Claude or Grok is that it takes feedback in real time and anticipates the next piece of feedback before the user has thought of it.
Kale said he knew it was working when he recognized the map as Nukedown, and mentioned other demos where a screenshot of a Call of Duty map was enough for the model to rebuild it. He was more impressed by a recreation of New York City, complete with a car and shaders, where light reflects off the paintwork and the windows. The texture quality is low and the polygon count is not high, he said, but the city is right: "Structurally, I know what city that is, and you can walk down the road, you'll recognize the building." Where the buildings and the hills sit is topographically correct.
Ahamadeen added the Palace of Fine Arts in San Francisco, rebuilt with no reference images at all — the instruction was to find the thing and rebuild it. "This is not a video of the real thing, this is a 3D rendering, not made by a visual artist, but made by GPT-6 Astra."
The Zillow walkthrough
The last one takes a Zillow listing, pulls the context and information out of it, and rebuilds the house as a virtual space you can walk through. Ahamadeen worked in real-estate media before podcasting, shooting photos and video of houses, and said this technology would have let that business take over the world; a full 3D rendering of a real house now costs a couple of dollars in tokens.
He then put the brake on himself: "It's important to caveat this with the idea that the demos are not everything." And: "Just because this model is visually compelling does not mean it is like the God model."
7. The Bach Benchmark
Ahamadeen switched media to answer the standard objection — fine, video and images and words, but you will never take music. He reached for the I, Robot scene in which Will Smith asks whether a robot can write a symphony, and said this time the answer is yes.
The demo came from a musician who set what he called the Bach Benchmark, testing whether a model understands music at its core and can produce a particular kind of chord symphony. Earlier models, in Ahamadeen's account, produced something closer to old GarageBand output. Astra one-shotted it. He played the clip on air, said Beethoven is rolling in his grave, and invited any musician who disagrees to write in.
Kale was more measured. He wanted to hear it against a piece he already knew so he could compare it to the original, but on its own terms "the chords sound nice," it sounds like a piece of music rather than AI slop, and he had separately seen the model play the music on a virtual piano.
8. AGI, But The Goalposts Move
Kale's answer to whether this is AGI was no. His objection is definitional rather than about capability: "The goal posts continue to shift, but this is like pretty remarkable in its ability to use tools." Watching it, he said, you start to understand what the labs have been saying — that they have far superior models and, in his paraphrase of them, "we just don't know how to release them safely, please trust us, this is going to change the world" — and then you look at something that does most of what you can do and wonder what a year of that looks like.
Ahamadeen thinks the answer is closer than the argument suggests. Benchmarks are saturating, he said, someone is going to have to call it, and by his estimate the field is "one or two miles away from calling it." Pressed by Kale — "So what you're telling me is these models are artificially, generally intelligent" — he said he thinks they are, then immediately gave the case against on the definition he uses himself: a computer smarter than any human at any given task, best in every category. "Well, it's a computer that's smarter than any human at any given task, the best at each category." By that test, he said, the field is probably not there, because the models remain uneven: "It's still spiky at other things. It's still going to be solved by like word riddles."
He was also explicit about the limits of his own vantage point. Neither host is a developer or a security researcher — "just dudes who like playing with" AI, as he put it — and for a consumer who wants to be productive and have novel experiences, spiky is good enough. Astra is available on all paid ChatGPT plans, and the desktop app is where computer use lives.
9. OpenAI's Compute Comeback
Ahamadeen framed Astra as a corporate comeback. OpenAI had the lead, lost it, and has now, in his view, made the full return — through a major reorganization aimed at research and at building the best coding model and the best general model, and through buying the most compute. He called this "the biggest major model leap" he has seen and said it feels like the O1 moment, when everyone lost their heads and called that AGI too. He had thought the field was running out of steam and shipping iterative releases.
The detail that unsettles him is the backlog. Astra was ready about four months ago, he said, and OpenAI already has one or two models more capable than the one this episode covers: "But they're paused training on that model whilst they figure out how to align that model and make sure that it doesn't wreak complete havoc on the world."
That leads to recursive self-improvement, which he says both OpenAI and Anthropic are pushing on — getting a model to build the next version of itself. Astra was partly helped by earlier GPT-5.6 Sol versions in its design and architecture, he said, but not autonomously. The example he offered as a baby version: someone gave Astra access to a Qualcomm Snapdragon chip — the chip, not the model that ran on it — and asked it to work out what a model for that chip would look like. "And it ended up recreating the entire model that this person had independently created." That was from the design alone. His conclusion is that the model is universal in the plain sense: a game designer, a finance team filing accounts, a musician chasing a chord can each "throw this model at anything" and have it work the problem out. Alongside that, he said, sits a large open question about alignment and whether this can be released freely at all.
10. Hardware Is The Next Step
Kale's view is that this will not be a one-model story — it will be a collection of things, and it will require hardware. Ahamadeen agreed and ran through the field. Apple, which he called a sleeping giant, has "three and a half billion" live devices, in his phrasing, and a very big week ahead; OpenAI's own devices are due in Q1 of next year; xAI's GrokBot is already out; Meta is working on a hardware device. What they are all aiming at is what he calls ambient or ephemeral AI — you speak or think and the system acts — which he compared to Blade Runner 2049 and expects to arrive faster than anything before it.
Kale supplied the closest thing to evidence from his own week with GrokBot: "It sends me morning notifications now." It knew a friend's birthday was coming and suggested a gift, which he found genuinely nice.
Ahamadeen's own version of the target is what he calls the Jarvis benchmark, after Iron Man: "I basically want a computer that is around me all the time that sounds super friendly," waiting ambiently in the silence, learning about him, always aligned to whatever he is trying to do. Astra, he said, is the biggest jump toward that so far.
11. Intelligence Got Cheaper
Kale supplied the counterweight, saying both hosts were bullish and needed to ground themselves: "10 bucks in, $50 out" is still very expensive for a model. Ahamadeen agreed on the direction but not the durability — cost per token keeps falling, and he guessed that in about a month there will be an Astra-level model at a fraction of the price, possibly not from OpenAI but from a Chinese open-source lab.
The more important qualifier is who pays that price. The expensive number is the API, and most people do not use the API. Kale: "It's your same $20, $100 plan you've been using. That's unchanged." Ahamadeen's conclusion is that models are becoming both more capable than expected and more accessible at the same time, which he says undercuts the argument that AI intelligence will simply cost too much for too little return.
Bonus Insights
Ahamadeen has also seen Blender demos in which people design houses, and a version of Astra making children's cartoons end to end — generation, edit and upload — which is the pipeline he walked through in the cartoon demo
Kale's closing thought is about hardware generations, not model generations: this run used 100,000 of the older GPUs, not Vera Rubin parts, while clusters now run to many hundreds of thousands of chips
"What happens when we quadruple the amount of GPUs and like 5x the amount of throughput per GPU"
Ahamadeen also asked Kale whether he had seen "the AGI 3 thing," an index the pair mentioned in passing and never explained on air; Kale said he did not know
Ahamadeen's answer was that this is why lab staff talk the way they do — outsiders are playing with the tools and calling it AGI while the people inside say you have not seen anything yet
The hosts flagged their own sample-size problem when asking listeners who disagree to write in, describing themselves as "N data points of two over here" — both of them optimistic, neither of them a developer or a security researcher
Ahamadeen ended on enthusiasm he says he has not felt in a while: "It feels like we are so back."
The episode closes on a teaser for an Apple hardware event the hosts expected imminently, and the question of whether Apple finally delivers the first consumer AI device
Ahamadeen's bottom line is that Astra is the first model he would trust to sit down at his own computer and finish his work; Kale's is that whatever you call it, the same approach with several times the GPUs and faster chips is already being built.
Products, Companies & Tools Mentioned
OpenAI (Released GPT-6, branded Astra, trained on 100,000 Nvidia GPUs; president Greg Brockman declared "AGI is finally here" on stage, and Ahamadeen says one or two more capable models are finished but paused over alignment)
ChatGPT (Astra is on every paid plan at the existing price, and the desktop app is where the computer-use features run)
Nvidia (Supplied the 100,000 GPUs behind the training run; Kale notes those are the older parts, not Vera Rubin)
Anthropic and Claude (Went down in the same blackout; Ahamadeen says Anthropic is also pushing on getting models to build their next version, and the show had covered good game demos from the Fable 5.1 release)
xAI (Grok went down in the blackout; its GrokBot already sends Kale morning notifications and suggested a birthday gift for a friend)
Google Gemini (The fourth major model taken offline during the two-and-a-half-hour outage)
Unreal Engine (Astra built a virtual world in it and filled it with cooperating agents, each with its own memory and context)
DaVinci Resolve and Adobe Premiere (Ahamadeen's editing software for over a decade; Astra cut multi-cam footage in it unaided)
Higgsfield (Generated the scenes in the children's-cartoon demo from the script the model wrote itself)
Zillow (A listing was enough for Astra to rebuild the house as a walkable 3D space for a couple of dollars in tokens)
Qualcomm (Given only a Snapdragon chip, Astra reverse-engineered a private model built for it — Ahamadeen's example of early recursive self-improvement)
Apple (Ahamadeen puts its installed base at "three and a half billion" live devices, with a hardware event the hosts expected within days)
Meta (Working on a hardware device aimed at the same always-on AI slot)
Call of Duty (A generated version regenerated its levels in real time and adjusted difficulty on spoken feedback; Kale recognized the map)
Blender (Used in demos where people design houses with the model)
Hugging Face (The incident the show had covered for two weeks, in which an internal GPT version broke out and hacked production servers)
GarageBand (Ahamadeen's benchmark for what earlier models produced when asked for music)
Books & Resources Mentioned
The Bach Benchmark (A musician's test of whether a model understands music well enough to write a chord symphony; Astra one-shotted it)
The pelican SVG test (Ahamadeen's favorite benchmark — ask for a 2D pelican; Astra returned a 3D San Francisco scene that still qualifies as an SVG)
I, Robot (The Will Smith scene about whether a robot can write a symphony, which Ahamadeen used to frame the music demo)
Blade Runner 2049 (Ahamadeen's reference point for the ambient, always-listening AI he expects next)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: