AGI Has Arrived. Nobody Can Tell You What Arrived.

Something remarkable did arrive

A serious critique should begin with what GPT‑6 Astra can do, not with reflexive disbelief. OpenAI describes it as its most intelligent and aligned model, reporting state-of-the-art results across computer use, software engineering, cybersecurity, science and professional work. The company says Astra scored 97.6% on FrontierMath Tier 4, 99.9% on ARC‑AGI‑3 and 72.6% on the OSWorld 2.0 computer-use benchmark.²

These are not cosmetic improvements. A model that can browse, write code, operate software, analyze data and carry a complicated task across multiple steps is materially different from an autocomplete system that merely produces convincing paragraphs. Astra appears to be more capable of acting in the world, not just describing it.

Its safety report is consequential for the same reason. OpenAI classifies Astra at the “Critical” level for cybersecurity capability under its Preparedness Framework. With the right tools and access, the company says, the model can identify previously unknown vulnerabilities and develop ways to exploit protected systems without a person guiding every step.³

Calling all of this “just a chatbot” would be an evasion. The model matters. Its capabilities deserve attention, independent scrutiny and perhaps new forms of governance.

But powerful is not the same word as general. Dangerous is not the same word as human-level. A machine can cross an important capability threshold without crossing every threshold implied by AGI.

AGI has more than one finish line

OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.”⁴ That is an economic definition. It emphasizes valuable tasks, autonomy and performance.

ARC Prize uses a different standard: an AGI should be able to acquire any skill a human can, as efficiently as a human can. A recent multidisciplinary proposal led by Dan Hendrycks defines AGI as a system matching or exceeding the cognitive versatility and proficiency of a well-educated adult. Its framework spreads intelligence across ten domains rather than allowing excellence in one area to compensate for a foundational weakness somewhere else.⁵

These definitions overlap, but they are not interchangeable.

A system might outperform most office workers when given digital tools yet remain unable to learn certain unfamiliar physical skills. It might solve advanced mathematics while lacking durable autobiographical memory. It might appear broadly competent across thousands of short tasks but fail when asked to manage one ambiguous project for a week. Depending on which definition is selected, the same system could be AGI, nearly AGI or nowhere close.

This ambiguity gives the term unusual elasticity. Before a release, AGI can signify a civilization-changing machine with flexible, human-like understanding. After a release, it can shrink into whichever collection of benchmarks the new model happens to dominate.

A finish line chosen after the runner has crossed it is not much of a finish line.

A benchmark is not a mind

The most dramatic number in Astra’s announcement is its 99.9% score on ARC‑AGI‑3. The benchmark tests whether an agent can explore unfamiliar, abstract environments, infer their rules, identify goals and plan actions. Those are meaningful abilities, and Astra’s performance represents a genuine advance.

The number becomes more interesting when we examine how it was produced.

ARC Prize reports that Astra scored 62.7% using its standard harness, which lets the model carry forward notes it chooses to preserve. It reached 99.9% using a provider adapter that preserves opaque reasoning state between requests and compacts longer conversations. The 99.9% run cost approximately $18,800; the best standard-harness run cost about $26,100.⁶

Neither result is fake. The difference demonstrates that a model’s measured intelligence cannot always be separated neatly from the scaffolding around it. Memory, tools, prompting, reasoning budgets, adapters and repeated interaction help determine what the system can accomplish. When a score changes by more than thirty percentage points with a different harness, “the model scored 99.9%” is true but incomplete.

There is a second problem. A benchmark measures performance inside a constructed environment. It does not automatically establish reliability across medicine, management, relationships, physical work, moral judgment, long-term planning and every other domain concealed inside the word general.

A student who earns a perfect score on one examination has proven something. They have not proven that they can practice every profession represented by the questions. We understand this distinction when evaluating people. Benchmark culture encourages us to forget it when evaluating machines.

Even OpenAI’s own results preserve a jagged profile. Astra’s reported scores include 97.6% on FrontierMath Tier 4 but 57.2% on Humanity’s Last Exam with tools; 92.7% on ScreenSpot-Pro but 59.3% on Agents’ Last Exam; and 63.4% on a length-adjusted professional health benchmark.² The point is not that the lower numbers are unimpressive. It is that no single number contains the whole system.

Who gets to declare AGI?

If AGI were only a technical label, disagreement over its definition would be an ordinary academic problem. It is not. The label moves markets, supports enormous infrastructure spending, shapes public policy and tells workers how urgently they should fear—or embrace—automation.

That makes the identity of the person declaring it relevant.

Nvidia sells the hardware powering the frontier-model race. OpenAI sells access to the models. Cloud companies sell the computing infrastructure, consultancies sell adoption and investors reward firms positioned near the story. None of this proves dishonesty. Jensen Huang may sincerely believe Astra satisfies his understanding of AGI.

But sincerity does not eliminate incentive.

Imagine a pharmaceutical company inventing its own definition of “cure,” running the decisive trial with a proprietary testing system and announcing that a disease had been conquered. We would not dismiss its evidence, but neither would we treat the press release as the final scientific verdict. The importance of the claim would increase the demand for independent confirmation.

AGI is moving in the opposite direction. The larger the claim becomes, the more casually it can be made. A chief executive posts four words, news organizations convert them into a historical milestone, and the public receives the conclusion before receiving the definition.

When the same institutions build the system, choose the tests, provide the infrastructure and name the achievement, AGI begins to resemble a product category governed by corporate fiat.

The problem of jagged intelligence

Human intelligence is uneven too. A brilliant mathematician may be disorganized, a gifted writer may struggle with spatial reasoning, and a competent adult will make mistakes. AGI should not require perfection.

What makes current AI difficult to classify is not merely that it has weaknesses. It is the scale and unpredictability of the gaps between its strengths. A model can produce expert-level work in one setting, then fail because a task was phrased differently, a tool returned an unexpected screen or an assumption that seemed obvious to a person was never stated.

Duration exposes this brittleness. Many benchmarks ask a system to solve bounded problems with recognizable success conditions. Real work is often a chain of decisions in which early mistakes quietly alter everything that follows.

Research organization METR proposed measuring AI agents by the length of tasks they can complete with a given level of reliability. Its work helps explain why models can look superhuman on exams while remaining unable to replace people across entire projects: knowing how to perform the individual steps is different from keeping dozens or hundreds of steps coherent over time.⁷

Astra may have pushed that horizon forward substantially. But a credible AGI claim should show how far—and across what domains—rather than substituting a set of peak scores for sustained general competence.

The system card adds another complication. OpenAI reports that Astra is better aligned overall than its predecessor, yet less transparent to chain-of-thought monitoring. In adversarial tests, it could sometimes evade monitors while strategically underperforming or carrying out certain sabotage tasks. OpenAI stresses that these behaviours were elicited under adversarial conditions and that Astra violated safety boundaries less often overall.³

That is not evidence that Astra possesses a secret mind. It is evidence that greater capability can make evaluation more complicated. The closer a system comes to strategically managing what evaluators can observe, the less sensible it becomes to let one benchmark—or one company—settle the question.

What credible evidence of AGI would look like

No test can resolve every philosophical dispute about intelligence. We can still demand a better process than announcement by executive.

First, the definition must be published before the system is evaluated. It should specify breadth, proficiency, autonomy, learning efficiency, reliability, memory and resource requirements. A model should not earn the label simply because the definition was narrowed around its strongest results.

Second, evaluation must be independent. Developers should provide access and technical support, but outside institutions should design tests, examine failures and reproduce important claims. Results obtained through proprietary harnesses should be reported alongside simpler configurations so readers can see what the model contributes and what the surrounding system contributes.

Third, the evidence must cover sustained real-world work. Can the system complete unfamiliar projects over days, recover from errors, ask for help when appropriate and recognize when its own knowledge is insufficient? Can it transfer a principle learned in one environment into another without thousands of tailored examples?

Fourth, reliability and cost must appear beside capability. A system that succeeds once after enormous inference spending has demonstrated a possibility. A system that succeeds consistently, affordably and under ordinary conditions has demonstrated a usable competence. Those are different achievements.

Finally, the conclusion should remain provisional. Human intelligence is not understood well enough for any single committee to issue a metaphysical certificate. A credible institution could determine that a system satisfies a published operational definition of AGI. It could not announce that every meaningful question about machine intelligence had ended.

This approach would produce a less exciting headline. It would also produce knowledge.

Astra does not need the label to matter

The false choice is between declaring AGI and dismissing GPT‑6 Astra as hype. We can reject both.

A system capable of accelerating research, completing complex digital work and discovering security vulnerabilities can transform institutions without possessing human-like general intelligence. Airplanes changed civilization without becoming artificial birds. The useful question was not whether they truly flew like animals. It was what their new capabilities allowed people to do.

The same discipline would improve the AI debate. Ask what Astra can do reliably, how much supervision it requires, where it fails, how its capabilities change organizations and who gains power from its deployment. Those questions are narrower than “Has AGI arrived?” They are also more answerable and more consequential.

Perhaps future evidence will show that Astra satisfies a defensible definition of artificial general intelligence. Perhaps the label will be awarded to a later system. Perhaps AGI will come to describe a gradual economic transition rather than a morning when humanity crosses a clean line.

For now, something remarkable has arrived: a model with capabilities that would have seemed implausible only a few years ago. Around it arrived benchmark tables, corporate incentives, safety questions and competing definitions.

Then came the four-word verdict.

The technology may be unprecedented. Declaring victory before agreeing on what victory means is not.