GPT-6 Astra Arrives With an AGI Claim, and World Models Have a Big Week
OpenAI says GPT-6 (codename Astra) opens the AGI era, backed by a 99.9% score on ARC-AGI-3 that comes with its own asterisk, plus new computer-use demos in Excel, Word, Blender, and Unreal Engine 5. Meanwhile, World Labs' Atlas, Runway's GWM Worlds 2, and a couple of MiniMax H3 hacks make it a loaded week for world models.
Two stories collided this week that don’t really belong in the same video except that they happened in the same seven days. First: OpenAI declared the AGI era, kind of, with GPT-6 — codename Astra. Second: the world-model side of AI video had an honest-to-god news week, capped by a Runway announcement that broke mid-recording. The full rundown’s in the video.
OpenAI declares the AGI era
GPT-6 — codename Astra — is OpenAI’s biggest training run to date: 100,000-plus GPUs at the Stargate site in Texas. Greg Brockman closed the press briefing calling Astra “a generational leap,” then dropped the mic with “welcome to the AGI era” — though he also said, worth noting, that the actual definition of AGI should be left up to users.
What got me almost as much as the claim was how it arrived. There was no big rollout moment — the first wave was embargoed press, and OpenAI’s own blog post on GPT-6 didn’t go up until a good two hours after that. The post itself is well put together — well enough that I half-joked Astra might have written it.
The headline number is a reported 99.9% on ARC-AGI-3, a benchmark built to be hard for AI, against Opus 5’s 30.2 and GPT-5.6 Sol’s 7.8. One asterisk worth keeping attached: Astra took the test with the help of OpenAI’s response API harness, so the number may be a little juiced. For what it’s worth, whatever the average human scores on the same test, the machines have clearly won — harness or not.
Astra as an agent: Excel, Word, Blender, Unreal
OpenAI is also billing GPT-6 as the world’s best computer-use model, and the demo reel does a lot of the convincing. A browser demo running through contact forms looked great — genuinely the kind of thing I want my agents doing — except it turned out to be sped up 4x. A separate Excel demo, shown at real speed, moved a lot slower. Worth remembering before you assume that agent speed is coming to your desktop as-is.
Other demos: reformatting a legal document (readable, though I’ve got no legal background to judge it against), and a design pass that visibly improved a before/after mockup. Playco showed a sizzle reel of Astra iterating on a Mario Kart-style racing game — polished, buzzword-heavy (“iterate quickly,” “less user intervention”), and I’d still like to know how long the underlying builds actually took. The demo that got me most was a house apparently designed in Blender, then run through Unreal Engine 5 independently, one tool feeding the next.
When you’ll get it, and whether it’s actually AGI
Astra isn’t in your model picker yet. It’s rolling out first to trusted testers through OpenAI’s Project Daybreak, with Pro, Plus, Enterprise, and API access promised “in the coming days.” It’s becoming the standard pattern for frontier launches — announce big, then wait — and it’s a little annoying, if still exciting.
As for the AGI question, I have my own doubts. Where this actually lands is going to come down to vibes as much as anything else, for me and probably everyone else watching — and it’s worth remembering OpenAI has an IPO to think about. What isn’t in doubt is that GPT-6 will be a genuinely strong model: GPT-5.6 Sol already is, and 6 is more than 5.6. Whether “strong” and “AGI” are the same word is a separate question.
The cost question: GPT-6 vs. Claude Fable 5.1
This is the part I care about. Anthropic put out Claude Fable 5.1 earlier this week, which I’d call the most capable model publicly available — a qualifier that matters, since GPT-6 isn’t public yet. Fable 5.1 is also expensive and burns through tokens fast: on day one, not paying close attention, I had it running on Ultra and burned through about 40% of my weekly quota on a single task.
So the real question isn’t AGI, it’s economics. If GPT-6 can do more for less, that’s a genuine win regardless of what you call it. If it costs three times what Fable runs, less so. I’ll know more once it actually ships. Until then, it’s back to yelling at Fable 5.1 to stop burning my tokens, and yelling at Opus to talk like a human.
World Labs’ Atlas
The biggest creative-side news this week was World Labs’ Atlas — a multimodal autoregressive diffusion transformer, which is a mouthful for: feed it an image, then step inside and explore it as a 3D world. The jump over their earlier Marble is image fidelity — Marble was impressive but prone to Gaussian-spike artifacts you’d usually end up cleaning up by hand in Nano Banana. Atlas’s headline feature is space-time simulation: feed it video, then walk around inside it.
It’s early access only right now. I’m talking to the World Labs team this week and expect access, so a full Atlas walkthrough is coming next week.
Runway’s Solaris, and GWM Worlds 2 breaking live
Runway is shipping Solaris — if that’s really a nod to the Tarkovsky film, well played. Runway calls it an interface world model: the obvious pitch is interactive, app-like websites, but watching it, I could also see an early version of a new interface for video inpainting and compositing, done live with no rendering step.
Then, while I was recording this, Runway announced GWM Worlds 2, its interactive world model — running at 720p, 24fps, with 48kHz audio, which most world models still skip entirely. The demo clip has two fully voiced characters trading directions to a clock tower. It landed too late for me to dig in properly, so I’ll circle back once I’ve spent real time with it — it’s the same real-time thread I’ve been pulling on since Lucy 2.5.
MiniMax H3, hacked into a world model
Two open-source projects are turning MiniMax H3 into something closer to a world model. First, H3-World, which repurposes the video model directly. I haven’t tested it yet, but I’d expect the usual world-model growing pains — permanence problems, where you turn away from an object and it’s not quite the same when you turn back — and I’m curious whether it avoids the “blobbiness” that creeps in the farther you wander from your starting point, the kind Google’s Genie 3 struggles with.
Second, Blendi built an open-source, Netflix-style wrapper around MiniMax H3 MAX’s real-time generation, called Unreel. The shows themselves are pretty basic. The more interesting part is what’s underneath: personalized, interactive, live-generated video. I don’t think Netflix CEO Ted Sarandos is losing sleep over it yet, but a smart-TV tile that generates a show around you, live, and lets you steer it, doesn’t feel far off. Not tomorrow — but given the pace of the last year or two, not absurd either.
AI summer never brought the heat everyone predicted, but between this and the real-time push, AI fall is shaping up to be a loaded season.
And on the question that matters more than the AGI label: I’m judging GPT-6 on the bill, not the benchmark. More for less than Fable 5.1 is the win. Three times the price is not. Verdict when it’s actually in my hands.