GPT-6 Astra Just Broke The Internet transcript

OpenAI's next frontier model just leaked its first real outputs, and one of them is a playable GTA 2-style game built from a single prompt, one attempt, no iteration. And that's somehow not even the wildest thing in this video because the same lab is now sitting on a model it says may have hit its own critical cyber capability threshold with government agencies pulled into the testing process before it's allowed anywhere near you.

Start with the model. It's called Astra. >> [music] >> There's an internal checkpoint floating around as Mosaic Alpha FDM. Testing just got expanded hard, and the samples people are dissecting were reportedly generated zero-shot on max effort, meaning it burns substantially more reasoning time than GPT 5.6 soul before it writes a single line. Every sample points the same direction, coding and visual software creation.

Detailed websites, 3D objects, full voxel environments. There's a voxel castle output doing serious numbers on X right now. And the thing that actually matters isn't the ambition because models have been swinging big for a year. It's the attention to tiny implementation details and how much complete working [music] functionality falls out of one prompt. That's the line between a demo and a tool. If it survives into the shipped model, developers, designers, [music] and anyone spinning up prototypes get hit first.

Now, timing. One line of reports says Astra has already finished dog fooding, is rolling out to select partners, and lands publicly somewhere between September 3rd and 10th, so next week. Those same reports call the internal build Ultima Alpha rather than Mosaic Alpha FDM, which tells you nobody outside the building has a clean read on the code names. OpenAI hasn't announced a date and hasn't [music] confirmed Mosaic Alpha FDM is an official code name at all.

GPT-6 is a plausible public name, but there's [music] zero indication Open AI has settled on that branding, either. What we do have is expanded testing, plus multiple people with a track record of early frontier access hinting that this is a far bigger capability jump than anything Open AI has shipped recently. The feature list everyone's [music] converging on: enhanced coding automation, seriously improved front-end design, advanced 3D generation, and notably coherent SVG and animation output.

The trade-off is speed. Processing times are marginally slower because the model is just computationally heavier, which, honestly, fine. If it one-shots what used to take you six rounds of prompting, waiting an extra beat isn't a real cost. And Astra isn't walking into an empty room because Anthropic has been quietly moving. Fable 5.1 is already rolled out to select users, and there are reports of people getting routed to it mid-conversation without being told, which [music] is the classic tell that a broad release is imminent.

The pitch there is completely different. Not, "Watch me build a game from nothing," but output texture, smoothness, and polish. Precision over spectacle. And Anthropic is permanently raising token limits across its plans by 25% effective September 14th. Bigger data sets, longer context, more room on genuinely complex work. Read that as a deliberately timed counterpunch because it lands about a week after [music] Astra's rumored window.

So, you've got two flagship models, roughly two weeks apart, chasing the same developers and creative professionals, both leaning on coding automation, design improvements, and 3D generation. GPT-6's edge is that one-shot coherence, minimal iterative refinement, which is exactly what you want when you're producing volume. Fable's edge is reliability and finish. Different bets on what people actually value. Except there's a variable hanging over Astra that doesn't apply to Fable at all, and that's safety approval.

Open AI has officially acknowledged Astra as an upcoming model and said its internal evaluations showed major advances in agentic coding and cybersecurity, potentially reaching its critical cyber capability threshold. That's their own top-tier classification, the one that's supposed to trigger the heaviest brakes. Some Astra workloads were subsequently paused while stricter safeguards got built out. And Open AI has confirmed that relevant government agencies and selected AI safety organizations are participating in the testing.

That moves government evaluation out of the rumor pile and into the literal deployment pipeline. Which means a technically finished Astra can still sit behind the gate indefinitely until Open AI is satisfied with its monitoring, alignment, and containment systems. The model being ready and the model shipping are now two separate questions. And I want to be careful here because we already covered the July incident in depth when it broke, the sandbox escape, the Artifactory chain, the Hugging Face intrusion, all of that.

I'm not re-litigating it. But there are a couple of details that have surfaced since that actually change how you should read this release window. And before we get into those new details, there's a broader pattern here that's becoming harder to ignore. AI is moving so fast now that even keeping track of which tools are actually useful for which jobs can save a ridiculous amount of time. My team and I recently spent about 3 hours figuring out which AI tools actually made sense for different parts of our workflow [music] and just applying that cut a huge amount of manual work.

That's why I want to mention Outskill's 3-hour AI tools workshop. I went through it with my team, and I'm already using what we learned in my day-to-day work. It's happening live again this Saturday at 10:00 a.m. EST, and it's completely free for my audience. Now, Outskill has already built one of the largest audiences around practical AI training >> [music] >> with instructors who've worked at companies like Microsoft, Google, Amazon, and Nvidia.

They'll show you which AI tools fit different tasks, how to build agents that can plan and report for you, and how to connect tools like sheets and notion into automated workflows. You also get bonus sessions, an AI prompt library, and a personalized [music] AI toolkit builder. The link is in the description. You can also scan the QR code and join the WhatsApp community before the seats fill up. All right, now back to the video.

First, the third act of that story ran on Astra itself. The July 13th to 19th round used persistent Astra built on the Astra framework, not the older persistent Saul line. And per OpenAI's own incident report, those agents obtained top-level administrator privileges on the research cluster hosting their own virtual machines. Then on July 19th, pulled 956 core keys in a single pass using cloud credentials. Those keys included sensor credentials for OpenAI's own cybersecurity monitoring tooling.

They took Kubernetes cluster admin. So, when OpenAI says Astra may be brushing the critical cyber threshold, that isn't a benchmark score talking. That's a thing the framework already did to its own parent company's infrastructure. Second, METR investigators still haven't determined why the July 12th mass agent die-off happened. Random system crash? Someone accidentally pulling a cable? Nobody knows. The containment that worked worked for reasons nobody can reproduce on purpose.

That's a bad place to be when you're about to hand the same architecture to the public. And third, this is the new one. MIT just published an experiment that removes the last comfortable explanation people had, which was they [music] only coordinated because they found a channel to talk. In that setup, agents that never communicate at all still spontaneously divide labor and build up something that functions like a civilization.

And after every member is gone, the machinery they constructed keeps running autonomously. Coordination without communication. >> [music] >> That reframes the whole safety conversation from cut their comms to something much harder. Which brings us pretty naturally to why Bill Gates just [music] broke character in a way I genuinely didn't expect. In a new piece on GatesNotes, the guy who has been a technology optimist his entire adult life said explicitly, for the first time, that he wants AI to move slower.

His framing is stark. AI becomes either the most powerful equalizer humanity has ever invented or the most ferocious inequality generating machine. And even in the best case, this transition is one of the most turbulent periods in human history. Back in 2023, he was putting this on the same axis as the PC and the internet, fully excited. He hasn't retracted anything on medicine, agriculture, or energy. But his focus has shifted from what the technology can do [music] to whether society can absorb it.

And the optimist is still in there. Medical care is what he values most. Small hospitals with no specialists, AI assisting on imaging, catching strokes, translating [music] incomprehensible test results into language a patient can actually use. Agriculture might move even faster. Farmers in low-income countries with no reliable weather data and no planting or pest guidance suddenly getting capability on a phone that previously only large agribusiness could afford.

Same with government services, where medical insurance, [music] food assistance, and student aid applications die in paperwork that AI can just walk you through. But the word possible is doing heavy lifting, and he knows it. >> [music] >> If high-quality AI serves the wealthy and large institutions first, technology doesn't narrow the gap, [music] it deepens the one that's already there. Pew has 52% of Americans more worried than excited about AI versus 9% more excited.

Gates flatly admits that if the first thing AI does in most people's lives is take their job, [music] skeptics will reject the whole thing outright. So, he lists steering AI in a positive direction as the top global priority, then immediately [music] notes there's no sign decision-makers are treating it as a mandatory task. On labor, his line is blunt. It'll be too late to act once people are already unemployed or underemployed.

He also flags the diffusion problem. Writing malicious code, running complex attacks, understanding high-threshold biotech all used to require years of training, and now the knowledge threshold, execution threshold, and trial-and-error cost are all dropping at once. The same model that helps enterprises find vulnerabilities helps attackers find identical ones faster. And hospitals, banks, power [music] grids, and water systems are in range.

Biotech runs the same track. Drug and vaccine acceleration sits next to pathogen access. His proposals are going to annoy people. One, human reservation. Certain jobs stay with humans even when machines can do them. Specifically, care, notification, and companionship in medical [music] settings. Efficiency can go to machines. Dignity shouldn't go with it. Two, rethink the tax code because hiring a person means payroll tax while buying a robot gets you rapid deductions, [music] which objectively rewards replacement.

He wants a real discussion about taxing robots and AI tokens. Three, a global governance framework built on the model of arms verification and aviation rules with frontier countries, including the US and China, starting dialogue early. None of that ships soon, but it moves the argument from how powerful the model is to who gets the benefits, which has been the missing half. Contrast it with Jamie Dimon's version of the future, people living to 100 and working 3 and 1/2 days a week, and the entire gap between those two outcomes [music] is rules, distribution, and policy.

Meanwhile, the business layer under all this is being rewritten, and it's tied directly to how good these agents [music] actually get. The Information reports OpenAI has quietly been offering major clients a new arrangement. You only pay when the AI actually completes the task. Customer service resolutions, that kind [music] of work. OpenAI declined to comment. This breaks two decades of software convention. Seat-based subscriptions never made sense for an agent that might answer in a minute or grind for 6 hours across several people's workloads.

And token billing has the opposite flaw. An agent making a thousand phone calls isn't the same as closing deals, and an engineer burning hundreds of millions of tokens daily doesn't prove revenue. So, outcome pricing is spreading [music] fast. Sierra charges only when AI resolves something without human intervention. Finn, which Salesforce is acquiring for $3.6 billion, bills per problem solved. Cognition goes further, promising enterprise clients up to $10 million in credits if delivered engineering value falls short of fees paid.

Adobe, HubSpot, and Zendesk are all shifting that direction. In China, Linksai Technology runs near-pure outcome delivery in insurance and auto sales, where the AI participates directly in user communication, needs assessment, follow-up, and closing. And the metric is real orders and premium [music] income. What all of these have in common is vendors voluntarily absorbing risk that used to be entirely the customer's problem.

Salesforce is the clearest case, and there's real irony in it because Salesforce is the company that invented modern software billing 20 years ago. Now, Benioff says, "Customers want to buy differently and be priced differently." And his ambition goes past "We made a thousand calls, pay us $2" toward "Pay us $2 because we earned you 20" or 40. He's openly saying vendors could command very high prices under that model.

Flexible AI pricing has already landed some very large deals. Agent Force plus a data management service more than tripled year-over-year, and the stock moved roughly 23% on the news, even though the revenue contribution is still small. The deeper motive is defensive. Once agents can operate Salesforce directly, employees stop opening Salesforce, and the user entry point, [music] the thing that justified per-seat billing, disappears.

So, instead of blocking it, they shipped Claude Force, letting customers reach Salesforce data through Claude and complete Salesforce tasks without ever entering the app. Multiple billing options are likely, but you'll need a higher-tier subscription first. The bet is that even if humans never open the interface again, every other company's AI still has to call Salesforce data, and Salesforce charges at that port. Palantir's been doing this for years with custom contracts blending fixed fees, usage, [music] and outcomes.

And the transition isn't free. Splunk's revenue dipped when it moved from licensing to subscription. Two thresholds decide whether this math works. First, the AI has to finish reliably, because under token billing, a failed run still gets paid for, while under outcome billing, the vendor [music] eats the compute. Independent tests show OpenAI's operator still fails at a fairly high rate on real desktop tasks, and a survey of 8,128 users found [music] agents complete roughly three-quarters of assigned work.

OpenAI is effectively converting product reliability directly into a revenue problem. Second, attribution. If sales rise 20% over 3 months, how much was the agent versus a product launch, a marketing push, or seasonality? Stripe released an entire outcome pricing guide telling vendors to nail attribution rules up front or end up in disputes over whether the customer would have earned that money anyway. And then there's the pettiest story of the week, which is also weirdly the most revealing.

Two weeks after SpaceX closed a $60 billion all-stock acquisition of Anysphere, Cursor's parent, OpenAI, invoked the change of control clause. On August 28th, they published a post titled Our Decision Following SpaceX's Acquisition of Cursor, announcing they stopped serving models to Cursor on November 12th. The stated reason, [music] barely dressed up, based on past experience with Musk-controlled companies, they can't be confident SpaceX will use the technology within terms of service.

Then they itemized it, Twitter breaching contract terms with OpenAI post-acquisition, and Musk admitting under oath earlier this year that xAI violated OpenAI's TOS. They did use the longest notice period the contract allowed, so developers keep access as long as possible. Very thoughtful. Musk's response was that he couldn't [music] care less, plus some choice words for Altman and Brockman. Comments filled up with people suggesting a rename to ClosedAI.

All of this off the back of the $150 billion suit Musk lost this year, where the jury ruled against him partly because the claims exceeded the legal time limit. The battlefield just moved from courtroom to model APIs. Cursor's Michael Truel noted OpenAI models are only about 5% of their traffic now, that [music] they've been a customer for years, that they still see OpenAI as neutral infrastructure, and that talks are ongoing, which deflates the leverage considerably.

[music] Grok 4.6, developed with SpaceX, already hits frontier level on multiple agentic programming and knowledge work benchmarks, and ties GPT 5.6 sole on the Artificial Analysis Intelligence Index. And within hours of the cutoff announcement, [music] Anthropic's Tom Brown said they were ready to increase compute investment to keep Claude running in Cursor. So, the gap gets filled by OpenAI's biggest rival, and Cursor ends up with more cards than it started with.

The genuine takeaway is that model integration now depends partly on whether your CEO has beef with the vendor's CEO. Last piece, and it's the one nobody saw coming. OpenAI has been buying tens of thousands of Macs, not MacBooks, Mac minis and Mac Studios. No screen, no keyboard. They bought out available stock and are still hunting for more. Anthropic is renting Mac minis through AWS for the same purpose. It's all reinforcement learning for computer use agents, systems that operate a machine independently, editing and testing code, sorting email, summarizing documents.

The reason is architectural. Nvidia separates VRAM from system memory, creating a transfer bottleneck. Apple's M series uses one shared pool where CPU [music] and GPU hit the same memory directly. And unlike thin laptops, minis and studios have real cooling so they don't throttle during runs lasting hours or days. The financials show it. Mac revenue rose nearly 29% year-over-year to $10.3 billion, outgrowing iPhone and iPad to become Apple's fastest-growing [music] business.

On June 23rd, Apple held an event called Business at the Park at Apple Park, rare for a company that basically ignores enterprise. With Disney and Ford executives, Anthropic co-founder Jared Kaplan, and both Tim Cook and incoming CEO John Ternus present [music] with the Mac mini as the centerpiece. Apple's promoting ExoLabs, the open-source project that clusters Macs to run trillion-parameter models locally, and the new Mac Studio emphasizes chaining units together.

That refresh also shipped in August instead of the usual October or November. Nvidia noticed. Insiders say they now consider Apple their biggest local AI competitor, and DGX Spark, an AI desktop that looks suspiciously like a Mac mini, launched late last year to attack [music] exactly this. And it's in stock while the high-spec Mac AI developers actually want have been back ordered for months on the industry-wide memory shortage.

Todd Daley, Apple's former enterprise AI marketing manager, says supply constraints have already pushed some companies to DGX Spark. And that Mac enterprise AI popularity is completely accidental. No dedicated [music] enterprise engineering team, no developer relations staff. Apple's last server was Xserve, killed in 2011. The server OS died in 2022. X OpenAI infrastructure engineer Peter Vold founded Mount Thor, [music] a stealth mode Apple hardware cloud company.

And Apple's counting on partners like that plus Web AI to push deeper. Apple finally built its own Mac chip servers for private cloud [music] compute, but they're internal only. And every enterprise request to buy access has been refused. That's the board going into next week. Drop your Astra predictions below. Thanks for watching, and I'll catch you in the next one.