OpenAI president believes Astra could mark the beginning of the AGI era, and the model does demonstrate a breadth of capabilities, using tools and performing complex tasks. But there is no agreed-upon test for general intelligence, and the existing documentation does not yet demonstrate independent learning, human-like reliability or full autonomy.
the mother General artificial intelligence Has it arrived yet? The question returned to the center of discussion with the launch GPT-6 Astra of OpenAIThe company's president, Greg Brockman, said that if we look back in a few years and ask when it was created... AGI, we may point to this period and even this model. However, its formulation was also probabilistic, not a declaration of compliance with an agreed scientific standard — because there is no such standard today.
The gap between the headline and the evidence is important. Astra It is undoubtedly a very advanced model, and perhaps the strongest candidate yet for describing a multi-purpose intelligence system. But “very advanced,” “multi-purpose,” and “artificial general intelligence” are not synonymous. To decide whether the threshold has been crossed, one must first clarify what the threshold is in the first place.
What is artificial general intelligence?
There is no single, agreed-upon scientific definition of AGI. The accepted definition in public discourse is a system capable of learning, inferring, and using knowledge across a wide range of domains at least at a human level. OpenAI has previously used a more economic definition: autonomous systems that outperform humans in most economically valuable work.
Other definitions emphasize adaptation to new tasks, learning from experience, transferring knowledge across domains, long-term planning, or successful operation in diverse environments. As early as 2007, Shane Legg and Marcus Hutter proposed measuring intelligence by an agent's ability to achieve goals in a wide range of environments. Their framework is important because it emphasizes breadth and adaptability, but it does not provide a single practical test that can be given to a model and labeled "passed."
This is where the main problem comes in: AGI is not currently an official rating like temperature or speed. Companies and researchers can use the same term and aim for very different levels of capability.
How does Astra approach AGI?
According to OpenAI's Official Model GuideAstra is the company's most intelligent model to date, delivering advanced performance across computing, browsing, software engineering, science and professional work. Equally important is its ability to perform multi-step processes that combine code, browser and professional software, rather than just a single textual response.
First, there is breadth. The same model can parse a document, search for information, write code, work with files, create a document, and perform actions on a computer interface. Such breadth is more reminiscent of a general knowledge worker than a narrow tool built for a single task, like facial recognition or playing chess.
Second, Astra is not just a conversation engine. According to Official specification page, it supports network and file searching, code execution, a command shell, computer usage, and connection to external tools. It has a context window of 1.05 million tokens and a maximum output of 128 tokens. The ability to combine large context with tools and multi-stage planning brings it closer to the character of a general agent.
Third, the model is able to remain coherent over longer tasks, receive new instructions mid-task, and change course without necessarily losing the big picture. These are not just convenient additions to the interface. The ability to adapt to new demands during a task is an essential component of the behavior we tend to call intelligent.
Fourth, OpenAI emphasizes achievements in science, mathematics, and professional work, and Brockman pointed to solving difficult mathematical problems and the ability to perform tasks that previously required experts. If a system can reliably contribute to a wide range of professions, it approaches the economic definition of AGI even without emulating full human capabilities.
How is Astra not yet a proven AGI?
The first barrier is reliability. Impressive success in a test or demonstration does not guarantee that the model will be equally successful in every new formulation, in a rare domain, or in a long workflow where early error may accumulate. A general system should not only succeed often; it must also identify what it does not know, deal with exceptions, and remain reliable as the environment changes.
The second barrier is continuous learning. Astra uses knowledge acquired through training and context provided by the user, but there is no published evidence that it independently updates its basic knowledge from each new experience as humans learn throughout their lives. A large context window and external memory can simulate continuity, but they are not the same as general and stable learning.
The third barrier is autonomy. Astra can select and run tools within an agent system, but OpenAI’s documentation makes it clear that the external application still executes the tool calls and manages pending jobs. That is, the practical capability does not belong to the model alone: it depends on permissions, tools, the host software, and the human operator.
The fourth barrier is dealing with the world in its entirety. The model accepts text and images, but according to its specifications, it does not directly accept audio or video. It also does not have a body, continuous senses, or physical experience in the world. One could argue that AGI does not require a robotic body, but anyone who defines general intelligence as equal to a human across the entire range of tasks cannot ignore this gap.
The fifth barrier is the lack of full independent validation. OpenAI reports top performance in several evaluation groups, but the public documentation does not make the phrase “AGI” a result that can be reproduced in an independent laboratory. Proprietary models, changing tests, and different definitions make it difficult to determine whether a new scientific status has been achieved and not just another engineering feat.
Is there a test that determines whether a system is AGI?
No. There are tests that check parts of the image, but no single test is accepted as an AGI seal.
Turing test Checks whether a person has difficulty distinguishing between a machine and a person. Language models They can excel at imitating conversation without demonstrating general learning, reliability, or the ability to function in the world. Therefore, it is no longer considered a sufficient test.
Knowledge, math, science, and programming tests test competency in important areas, but they may also measure familiarity with similar questions, focused practice, or computational power. ARC-AGI Designed to test abstraction and adaptation to new puzzles, it is particularly useful for measuring a component of fluid intelligence; even a high score on it does not prove general competence in medicine, long-term planning, social relationships, or the physical world.
Agent tests test whether a system can complete a long task using a computer and tools. They bring the assessment closer to the real world, but are still highly dependent on the test environment, permissions, and the amount of assistance the system receives.
A more convincing claim about AGI will therefore require a bundle of evidence: independent and consistent success in new tasks across a range of domains, learning from experience, transferring knowledge across domains, long-term planning, error detection, and resilience to unplanned conditions. It is desirable that independent groups can also reproduce the results.
So what is the verdict?
Astra may be the closest system yet to what many people call functional general artificial intelligence: it is multi-disciplinary, operates tools, performs complex tasks, and adapts to changing instructions. To a user looking at the result rather than the architecture, it can sometimes seem like a very capable human counterpart.
Still, there is currently no solid scientific basis for the claim that Astra has “achieved AGI.” Even in its official documentation, OpenAI calls it its most capable and intelligent model, but does not give it an unequivocal AGI label. The report on Brockman's remarks Presents a possible historical assessment — perhaps in the future we will look back and see this moment as the beginning of the era — and not the result of a decisive test. The same report even quoted the company's CEO Sam Altman as calling AGI a poorly defined term and even an irrelevant marketing term.
So the precise wording is that Astra greatly narrows the gap to some of the definitions of artificial general intelligence, and perhaps forces us to update the meaning of the term. It still does not prove that the gap has been completely closed.
Questions and Answers
Has OpenAI officially declared Astra to be AGI?
Company president Greg Brockman has said he believes Astra may mark the beginning of the AGI era. However, the official technical documentation describes it as OpenAI’s most capable and intelligent model, but does not explicitly classify it as artificial general intelligence.
Did Astra pass the AGI test?
There is no single, agreed-upon AGI test. There are tests for abstraction, knowledge, programming, computer use, and planning, and each of them measures only a part of the overall ability.
Is Astra smarter than humans?
In some tasks, it may outperform most humans in speed, knowledge, or accuracy. This does not mean that it outperforms humans in every area or is equally reliable in every new situation.
Does AGI require consciousness?
Not necessarily. Most practical definitions deal with the ability to perform tasks, learn, and adapt, rather than with internal experience or consciousness. There is currently no scientific evidence that astra is conscious.
What will it take to be convinced that AGI has been achieved?
Independent, reproducible evidence of human or superhuman performance on a wide range of novel tasks, over time, with a low failure rate, learning from experience, and adaptation to situations that did not arise in training.
The scientific article
More on the subject on the science website
- A breakthrough in the field of artificial intelligence: GPT O3 is approaching a general level
- Does the Turing test meet the GPT test?
- New study: Language models are still not reliable enough for core software development tasks
- Why OpenAI's solution to AI "hallucinations" will eliminate GPT chat tomorrow morning
- General artificial intelligence is almost here
