The AGI Nobody Can Define

What Astra Actually Did

GPT-6 Astra shipped on September 3, 2026, first to customers in OpenAI's Daybreak cybersecurity program, then more broadly through the API within the week.¹ It is, by any measure OpenAI has published, the most capable model the company has released. On FrontierMath Tier 4 — a benchmark built specifically to resist memorization, using problems research mathematicians rate as genuinely hard — Astra scored 97.6%, up from 83% on the previous flagship model, GPT-5.6 Sol. On ARC-AGI-3, a test designed to measure reasoning that can't be pattern-matched from training data, it hit 99.9% under OpenAI's own test conditions. On GPQA Diamond, the graduate-level science benchmark, it scored 96%.²

The cybersecurity numbers are the ones that got Astra classified as “critical” under OpenAI's Preparedness Framework — the first model to earn that label. It scored 100% on ExploitBench and, during evaluation, discovered two previously unknown zero-day vulnerabilities across real browsers and operating systems on its own.² And in a report OpenAI released weeks before the launch, the company said Astra-family models had produced verified solutions to ten mathematics problems that had gone unsolved for at least a decade — spanning high-dimensional geometry, coding theory, quantum complexity, lattice cryptography, and extremal combinatorics. One result settled a genuinely open question in group theory: the existence of non-sofic groups.³

Brockman called it “the most intelligent and most aligned model yet” and, on stage, said something close to a personal verdict: “I do think we're there.”¹ He wasn't hedging for effect. Sam Altman had said weeks earlier that he expected an internal system qualifying as AGI “by the end of the year” — and OpenAI's chief research officer, Mark Chen, had put a number on it: roughly 80% of the way.⁵

The cybersecurity capability is why Astra became the first OpenAI model classified as “critical” under the company's own Preparedness Framework — the internal risk tier reserved for systems judged capable of meaningfully assisting someone in building a cyberweapon, not just answering questions about one. Human red-teamers confirmed Astra could identify genuinely novel zero-day vulnerabilities across real browsers and operating systems, unassisted.¹ OpenAI's research VP, Amelia Glaese, tied that capability directly to the trust problem sitting underneath the whole launch: “When models can do more things autonomously, we have to be able to trust them more.”⁴ It's a tidy sentence that doubles as an admission — the more autonomous the system, the more the humans supervising it are working on faith.

Why It's Fast — and Why That's the Problem

Part of what makes Astra's numbers possible is a genuine architectural change, not just more training compute. Astra uses what researchers call recurrent depth, sometimes described as a “looped transformer.” A conventional large language model pushes each token through a fixed stack of layers once, in sequence. Astra instead feeds information through a single block repeatedly, reusing the same weights multiple times rather than writing every intermediate step out to a place a person could read.⁶

The payoff is real: studies of the approach show it can match standard transformer performance while using 50% to 90% less compute.⁶ That is not a marginal efficiency gain. It changes what a lab can afford to run at scale, and it's a large part of why OpenAI's own cost-per-task figures for Astra look better than its raw per-token price — roughly 57% lower API cost on coding tasks despite a 2.5x sticker-price increase on tokens.²

That's the catch researchers are worried about. Because Astra's reasoning loops internally rather than narrating itself in plain language, its intermediate “thinking” isn't fully expressed as readable text the way earlier reasoning models' chain-of-thought was. Researchers have started calling this opacity “neuralese.” OpenAI's own chief scientist, Jakub Pachocki, acknowledged the trade-off directly: “as model capabilities are increasing, monitorability is getting more challenging.”¹ Steven Adler, a former OpenAI safety researcher, was blunter: “OpenAI seems to be violating one of the few redlines that exists in the AI industry.”⁷ Peter Wildeford of the AI Policy Network called the move “potentially very concerning” and “potentially reckless,” pointing out that chain-of-thought transparency had been central to investigating an OpenAI security incident just two months earlier.⁷

There's a second, quieter implication buried in the efficiency story. OpenAI has been explicit that Astra isn't just a product — it's a step toward what the company calls a “research-intern-level” automated AI system, a milestone it targeted for September 2026, on the way to a fully autonomous AI researcher it has pencilled in for March 2028.³ Astra is described internally as able to coordinate multiple agents over extended stretches — hours, sometimes days — on a single hard problem. Cheaper compute per unit of reasoning is exactly the resource that determines how much of that research loop a lab can actually run. The efficiency gain isn't incidental to the self-improvement roadmap. It's the budget line that makes the roadmap affordable.

The Argument Over the Word

Here is the part of the story that Astra's benchmark scores don't capture: OpenAI's own executives can't quite agree on what they just announced. Brockman, at the same briefing where he said “welcome to the AGI era,” also admitted “there's no clearly defined AGI moment” — that the transition “has been more gradual than expected” compared with what the founding team pictured in 2015.⁸ When directly asked whether Astra constitutes AGI, he reframed the term itself, noting that the old contractual trigger — a specific definition written into OpenAI's agreement with Microsoft — no longer applies, and describing AGI now as more of a “mission concept or spiritual concept” than a technical bar to clear.¹ Altman's own framing, weeks earlier, was even more candid about the sleight of hand: OpenAI would have AGI by the end of 2026, he said, “if you accept his definition.”⁵

Not everyone is buying the reframe. Gary Marcus, the cognitive scientist who has been arguing against premature AGI declarations since long before Astra existed, published his response within hours of the launch. His core objection is what he calls the “fallacy of composition”: the assumption that excellence in one domain — here, mathematics that happens to be unusually easy to verify and to generate synthetic training data for — implies excellence everywhere. “Expertise in one domain does not at all guarantee expertise in all or even most domains,” he wrote.⁹ His verdict was direct: “I see no reason whatsoever to think Astra is AGI let alone ASI.”⁹ Following mathematician Ernie Davis's analysis, Marcus also flagged a quieter methodological gap in OpenAI's math claims: the company hasn't disclosed how many conjectures Astra attempted versus how many it actually solved, or what the computational and human-labour cost of each success was.⁹

Anthropic, OpenAI's closest rival, has taken a different rhetorical path to a similar place. Dario Amodei has talked about AGI-level capability arriving within “one to three years” rather than declaring an era on a specific Tuesday — a hedge, not a declaration.¹⁰ And Anthropic's own recent milestone — Claude writing more than 80% of the company's production code — was announced as an engineering statistic, not a philosophical one. Anthropic has also kept its own most capable system, an internal model reported to score 97.6% on a mathematics olympiad and find zero-day vulnerabilities unassisted, out of public release entirely, restricted to vetted cybersecurity defenders. Two labs, two very different appetites for saying the word out loud, and neither one waiting for the other to blink first.

The Research Loop Is the Real Story

Strip away the word “AGI” entirely and there's still something worth taking seriously in Astra's launch, and it's the piece closest to the ground: OpenAI has said, on the record, that it is building toward a “research-intern-level” automated AI system this year and a fully autonomous AI researcher by March 2028.³ That is not marketing language about a spiritual concept. It's a stated engineering roadmap, and Astra — a system explicitly designed to coordinate multiple agents across hours or days of sustained problem-solving — is presented internally as a milestone on that specific path, not a general-purpose chatbot that happens to be very good at trivia.

OpenAI isn't alone in formalizing this. A July 2026 paper out of Tsinghua University and the startup Frontis.AI, titled plainly “AI4AI,” describes training a model whose job is explicitly to improve the process that builds AI models — a concrete, measurable step toward recursive self-improvement using machine-learning engineering itself as the test bed. The system, Frontis-MA1, improved a 35-billion-parameter base model's problem-solving score from 39.4% to 60.6% purely by getting better at proposing and debugging its own training code, in a loop the researchers describe with a line that could double as Astra's mission statement: “each improved system further improves the process that produces its successors.”¹¹

That is the sentence to hold onto, more than “AGI era.” Whatever Astra is or isn't by whichever definition a company chooses this quarter, the operational fact underneath the announcement is straightforward and doesn't depend on anyone agreeing on a label: the frontier labs are no longer just building models that answer questions. They are building models whose explicit job description includes accelerating the research and development of the model that replaces them. Brockman's four words were a marketing choice. The research-intern-to-autonomous-researcher roadmap is the actual news — and it was true the day before he said them, too.