Dave's first impression of OpenAI's new model was not a score. It was a picture of his own screen. This is the first class of models, he said, where as you are talking to it, it is "showing you screenshots of your own laptop saying, is this what you wanted?" — and the image simply appears in the text stream, with no third-party component to install. Before, he said, the model would come back with long-winded technical explanations you could spend twenty minutes decoding. Now it shows you two options and asks which one you meant.
That moment sits oddly beside the leaderboards discussed on the same episode of Moonshots with Peter Diamandis. Depending on which measuring stick the panel put on screen, GPT-6 — which OpenAI calls Astra — was either the most capable model in the world or the third most capable model in the world. Working through that contradiction took up most of the discussion, and it led the panel away from "which model is smartest" toward a narrower question: what was this model actually built to do?
What OpenAI announced
Peter read from OpenAI's release page, which describes Astra as state of the art on computer use, browsing, software engineering, cybersecurity, science and professional work. The numbers he relayed: near-saturation of FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9%, ExploitBench at 100%, and a hallucination rate — the rate at which a model states things that are not true — falling by nearly half while accuracy rose. OpenAI's own release table puts the math figure at 97.6% on Tier 4 v2, and shows how much the security results depend on which tasks are used: 100% on ExploitBench, but 39% on its June–August 2026 subset and 42.4% on ExploitGym.
In a Bloomberg TV interview played on the episode, Sam Altman described the change in everyday terms. If you have an idea and want to work interactively with an AI to get a whole piece of software built, he said, "this is the first model to me where I could sort of tell someone like, just give it a try. There's a good chance that'll work." He said he had watched people make computer games, do DIY electrical engineering projects and run complex simulations for their own scientific work, and expected users to start throwing harder tasks at it as trust built up.
Three rulers, three answers
Alex walked the panel through the charts. The first was the Epoch Capabilities Index, maintained by Epoch AI, which he described as one possible suite that leans heavily into math and other technical fields. Epoch's documentation explains the trick behind it: because many models take many tests, the index can infer how hard each benchmark is and place easy and hard tests on a single scale, so comparisons survive the saturation of individual benchmarks. It is an inferred capability score, not a percentage of questions answered. On that scale, Alex said, Astra is number one in the world, ahead of Claude Fable 5, on a trend line that has been "perfectly predictable going back years." One component is FrontierMath Tier 4 v2 — version two, he noted, because version one contained incorrect answers that AI itself had to correct. "Math is thoroughly cooked at this point."
The second ruler gave a different answer. On Artificial Analysis's intelligence index, which Alex called "maybe a little bit more focused on broad, economically valuable activities," Astra is not first. Anthropic's Claude Fable 5.1, released days earlier, holds the top spot; Muse Spark is the second vendor family; GPT-6 is a close third. Artificial Analysis's methodology combines ten evaluations with weights of 30% agents, 20% coding, 20% scientific reasoning and 30% general capabilities — including hundreds of SaaS automation tasks graded on objective completion, with no credit after a guardrail violation. The suite is mostly English and text-based, and its version history changes tasks, graders and weights, so a comparison only means something alongside the index version and the model's reasoning settings.
Plotted against cost per task, Alex said, Astra at maximum reasoning "isn't even at the frontier," sitting just below Claude Opus 5, with Chinese and Meta models nearby. Then came the third chart, and the reason he thought the first two were arguing past each other: capability against output tokens per task. There, he said, the frontier "is just dominated by GPT6."
Fewer tokens, because the model is driving
Output tokens are the units of text a model produces, including the reasoning it writes out before answering. Fewer tokens for the same job means lower cost, less waiting and a different feel to the product.
Alex's reading — offered as inference, since he said he did not yet have access to Astra — is that OpenAI deliberately optimized for this through architectural choices and targeted applications, because it had computer-use assistants in mind: software that operates a desktop or browser rather than calling an API. To be good at that, he said, a model needs to understand what is on the screen, needs native multimodality, and needs to parse video, images and screenshots "in a really tight interactive low latency loop." And it is "slower to do all of your reasoning out in chain of thought versus as a feed forward pass of your transformer." He framed this as OpenAI finally internalizing a story Anthropic told first: when Anthropic leapfrogged OpenAI on revenue, he said, it did so on high-value enterprise code generation and on computer use, where Claude Code could reach all the tools in a user's local desktop environment. Astra, he said, looks like "a next generation frontier model that's designed with CUA from the ground up."
OpenAI's release video, played on the episode, is a tour of exactly that. A user asks for a yellow circle, turns it into a rocket window, then says "now make it a 3D model and blender" and the model opens Blender. Elsewhere it builds a retail presentation for next season's rainwear, creates an eBay listing for an orange flea-market table using a photo from the user's downloads folder and adds a note that it is slightly damaged, drafts a licensing agreement template and tightens the limitation of liability in the licensor's favor, looks for a tennis court in the Lower Haight, and exports an STL file for a 3D printer.
What ARC-AGI-3 was supposed to prove
The headline benchmark deserves its own explanation. ARC-AGI-3 does not ask questions; it drops an agent into an unfamiliar little animated world with no instructions and expects it to work out the goals and rules by acting. Alex described these as pixelated block worlds — some look like Tetris — where the challenge is to learn "the mini physics" of a game you have never seen, in real time. Scoring counts efficiency, not just completion: the benchmark treats 100% as beating every game about as efficiently as humans do, so brute-force trial and error is not the same achievement as learning the rules economically. ARC also bars external harnesses — the scaffolding code that wraps a model with tools and retry loops — because it wants to measure the baseline model.
Earlier models scored under 10%. Astra, Alex said, "just runs away with the game," at either near 100% or 60-plus percent depending on how it is measured. He offered two readings and did not choose between them: either the model is genuinely excellent at program synthesis, or OpenAI "has just done a really aggressive job of distilling the harness code that everyone else is using to beat ARC AGI3 back into the baseline model" — that is, folding the scaffolding's tricks into the model's own weights.
Dave thought the significance ran deeper than the leaderboard. ARC-AGI-3 was meant to be solvable by a bright 12- or 13-year-old and to stand for years as evidence that AI was not on the right path, he said, "and it just got obliterated so quickly."
The architectural guess, and what was not disclosed
Asked what changed inside the model, Alex was explicit that he was speculating from public comments by OpenAI leaders and analyses by others. His best guess is recurrence: looped transformers, where a single transformer is stacked on itself with the same weights — weight tying — and run around twice rather than once. He connected it to Chinese labs injecting recurrence elsewhere, such as Kimi's linear attention mechanism, and to academic work using looped transformers to get strong ARC results from very small models. If true, he said, it might be the start of "a new scaling law, which is depth scaling, which we've never seen before." None of this appears in OpenAI's announcement; it is a hypothesis about an undisclosed design.
Emad called Astra "the first non-benchmaxed model" — his shorthand for one whose scores were not tuned to the tests — and pointed at the same split Alex had charted: saturating ARC-AGI-3 while lagging on Artificial Analysis. He cited Greg Brockman saying this was OpenAI's first new pre-trained model since GPT-4o, which Emad said he found hard to believe, and estimated a training run on 100,000 next-generation chips costing on the order of a billion dollars, against Chinese pre-training runs he put nearer $10 million. A later remark on the panel added a caveat to everything users will experience: a model trained on 100,000 chips would need many times that number to serve, so what ships is a smaller, distilled version that runs in real time — "much smaller and not as smart" than the system that was trained.
The demos the panel wanted instead
What the panel did not like was the marketing. Building a 1970s-style video game, Dave said, badly undersells what is in the model. The benchmarks have the same problem: the half-life of each new test keeps shrinking, and what is needed now are tasks with consequences — solving diseases, designing a city, fixing urban traffic.
Salim took the complaint to ChatGPT, typed in that the Astra video was "pretty pathetic," and read out the three alternatives it proposed: plant a vulnerability in a large unfamiliar open-source codebase in a sandbox and see if the model can map the architecture, find it and fix it; create a synthetic $500 million manufacturing company with an ERP system and CRM records and give the model twenty minutes to produce a P&L and a recovery plan; and run a live disaster-response command center from simulated earthquake imagery and traffic-camera video, verifying conflicting reports and predicting which hospital roads would become overwhelmed.
That is the gap the panel kept circling. The demos show one assistant helping one person, in a straight line, while a panelist argued the real prize is optimizing supply-chain logistics or managing a million-person organization — implications "massively bigger than building a rocket video game in your basement." Someone put the limit plainly: "it'll run the computer, but it won't run an entire organization. Like it won't make a billion dollars for me." The reply was that the only restriction now is imagination; Peter told listeners to take off the shackles of what they think they are able to do and then go ten times bigger. The panel also noticed where OpenAI's video begins — in a paint program — and read the whole sequence as a bid on the desktop itself: not an application sitting on the operating system, but something behaving like one.