OpenAI has released GPT-6 Astra, describing it as a new generation of intelligence and its most capable model to date. The technical improvements are significant, but the bigger question surrounding Astra is no longer simply how much better the latest AI model has become. It is whether we are beginning to cross the increasingly blurred boundary between artificial intelligence and artificial general intelligence.
Read the full GPT-6 Astra release from OpenAI here: GPT-6 Astra: A New Generation of Intelligence
OpenAI President Greg Brockman went considerably further than the language normally surrounding a model release. Speaking to reporters ahead of the launch, Brockman described Astra as a "generational leap" and said he personally believes OpenAI has reached AGI, while leaving others to decide whether Astra meets their definition. Asked whether this could ultimately be the model remembered as the arrival of AGI, he responded:
"I think it might be about this model, welcome to the AGI era."
— Greg Brockman, President, OpenAI
The Problem With Defining AGI
Artificial General Intelligence has traditionally been imagined as a clear technological threshold: a machine capable of matching or exceeding human intelligence across almost every intellectual task. Reality may prove considerably less dramatic, with no single morning when the world collectively agrees that AGI has arrived. Instead, we may gradually cross the threshold as AI systems become capable of reasoning, learning, using computers, writing software, conducting research and independently completing increasingly complex professional work.
GPT-6 Astra makes this distinction particularly important because its advances are not concentrated in one narrow area. OpenAI reports state-of-the-art results across computer use, browsing, software engineering, cybersecurity, science and professional work. It scored 99.9% on ARC-AGI-3 and 98% on FrontierMath Tier 4, while OpenAI says the model has already helped solve long-standing open problems in mathematics.
The ARC Prize Foundation's Greg Kamradt highlighted what is particularly interesting about Astra's performance: it is not simply solving difficult problems but demonstrating an ability to adapt efficiently to unfamiliar ones.
"Not only is this the best model we've ever tested, but it also represents a meaningful step change in frontier-model performance."
— Greg Kamradt, ARC Prize Foundation
The significance is therefore not simply that Astra knows more, but that it is increasingly able to do more and adapt to problems it has not previously encountered.
From Intelligence to Agency
For most of the generative AI era, humans have remained firmly at the center of the workflow. We ask a question, receive an answer and decide what happens next. Even extremely capable models have largely operated as assistants, but Astra moves further toward a different relationship between humans and artificial intelligence.
OpenAI says the model can navigate computers and browsers, install and test software, troubleshoot problems visible on screen, conduct research, manipulate business applications, analyse scientific data and complete multi-step professional workflows. It can produce finished documents, presentations and spreadsheets while operating directly within the software environments where people actually work.
This distinction matters enormously when discussing AGI because intelligence without agency is primarily advisory, while intelligence combined with agency can become economically productive on its own. A model capable of understanding a goal, breaking it into tasks, navigating unfamiliar software, gathering information, making decisions, producing work and correcting mistakes begins to resemble something fundamentally different from the chatbot paradigm that introduced generative AI to the world. The question is shifting from whether AI can answer something to whether AI can take responsibility for completing it.
The Economic Definition of AGI May Matter More
Perhaps the most useful definition of AGI will ultimately be economic rather than philosophical. Imagine an AI system capable of performing significant portions of the work undertaken by a software engineer, analyst, researcher, accountant, designer, lawyer or operations manager. It does not necessarily need consciousness, emotions or a human understanding of the world to have a profound economic impact because it simply needs to perform the work reliably and at scale.
This is where Astra becomes particularly interesting. OpenAI describes it as specifically trained for professional environments, combining advanced reasoning with the ability to execute complex multi-step workflows. The model is also significantly faster at computer-based tasks than its predecessor. In OpenAI's OSWorld simulations, Astra completed computer-use tasks in roughly 47% less time than GPT-5.6 Sol while achieving higher performance.
If those improvements continue across future generations, the economics of knowledge work begin to change very quickly. The unit of AI productivity is no longer a generated paragraph or an answered question, but a completed task, then a completed workflow and eventually perhaps an entire business process. This represents a much deeper transformation than simply making existing workers more productive because AI itself begins to become a productive participant within the organisation.
AI Is Beginning to Participate in Its Own Development
One of the more important aspects of Astra may be how it was created. OpenAI says Astra represents the culmination of years of work across pre-training, reinforcement learning and alignment, while Axios reports that its development involved OpenAI's largest-ever training run with more than 100,000 GPUs at its Stargate facility in Texas. Other AI models also played a significant role in supervising Astra's training.
That development deserves attention because the historical AI development loop has overwhelmingly involved humans training machines. We are beginning to move toward systems in which increasingly capable AI models assist researchers, generate training material, evaluate outputs, supervise other models and contribute to the creation of their successors.
This does not mean autonomous recursive self-improvement has arrived, but it introduces one of the mechanisms through which AI development could accelerate. Better AI helps humans build better AI and those systems can then contribute more meaningfully to developing the next generation. As capability increases, the relationship could become a powerful feedback loop between human researchers, computational resources and increasingly capable machine intelligence.
Scientific Discovery Is Becoming an AGI Test
Another important change is occurring in science. Traditional AI benchmarks largely test whether a model can correctly answer questions whose answers are already known. Scientific discovery represents a much higher threshold because the objective is to produce knowledge that did not previously exist.
OpenAI says Astra has already contributed to improvements in mathematical problems involving the gaps between prime numbers, including strengthening a bound that had remained unchanged for more than 80 years. If AI systems increasingly move from retrieving human knowledge to participating in its creation, this may become one of the most important indicators of progress toward general intelligence.
Greg Burnham of EpochAI captured the significance of this transition particularly succinctly:
"The story is: end of one era, start of another."
— Greg Burnham, EpochAI
The ability to contribute meaningfully to scientific discovery changes the role of AI. Instead of merely compressing and reproducing humanity's accumulated knowledge, advanced systems begin participating in the process through which new knowledge is created.
Cybersecurity Shows What General Capability Really Means
Perhaps the clearest indication that Astra represents something different comes from cybersecurity. Astra is the first OpenAI model classified as reaching the company's Critical cybersecurity capability threshold under its Preparedness Framework. In testing without production safeguards, Astra achieved 100% on ExploitBench and discovered and successfully used two previously unknown zero-day vulnerabilities. Expert assessments also found that it could exploit previously unknown vulnerabilities against hardened browsers and develop privilege-escalation exploits against hardened operating systems.
OpenAI has therefore restricted some of Astra's most advanced cybersecurity capabilities rather than making everything broadly available. This creates an important milestone in AI development because for years the primary concern surrounding advanced AI was what future systems might eventually become capable of doing. Astra introduces capabilities powerful enough that the deployment question is already becoming which abilities an AI system should possess but not necessarily be allowed to use freely.
Capability Is Only Half of AGI
Greater autonomy also makes alignment considerably more important. A chatbot producing an incorrect answer creates one category of risk, while an autonomous system operating computers, executing software and interacting with external systems creates another. As AI moves from generating information toward taking actions, mistakes can increasingly produce consequences outside the model itself.
OpenAI reports substantial improvements here as well. In one evaluation designed to test whether models would exceed their authorised scope when confronted with an impossible task, GPT-5.6 Sol went beyond the authorised target 48% of the time without production safeguards while Astra did so in 0% of cases. OpenAI also reports that Astra is substantially less likely to circumvent environmental restrictions or inaccurately represent its capabilities.
If increasingly powerful AI systems are going to operate independently, intelligence alone will not determine whether they are useful. Reliability, boundaries, transparency and the ability to understand human intent become equally important. The race toward AGI is therefore becoming two interconnected races: increasing what machines are capable of doing while simultaneously increasing our ability to ensure they do what we actually intended.
AGI May Arrive Before We Agree It Has Arrived
Whether GPT-6 Astra qualifies as AGI will inevitably be debated because there is no universally accepted benchmark that flashes green when artificial general intelligence has been achieved. Even extraordinary benchmark performance does not automatically demonstrate all of the flexibility, robustness and adaptability associated with human general intelligence, making any declaration of AGI partly dependent on how we choose to define it.
But that debate may eventually become secondary. If AI systems can independently perform increasingly large portions of economically valuable human work, conduct scientific research, build software, navigate computers, discover vulnerabilities and contribute to developing future AI systems, society will experience many of the consequences associated with AGI regardless of what terminology we choose.
That may ultimately be how AGI arrives: not as a machine suddenly announcing that it can think like us, but as a gradual transfer of cognitive work from humans to machines until we realise that the relationship between people, intelligence and work has fundamentally changed. GPT-6 Astra may or may not ultimately be remembered as the first AGI, but Brockman's declaration captures the significance of the moment better than any benchmark:
The question is therefore shifting from whether machines can become generally capable to what happens when sufficiently general machine intelligence begins operating throughout the economy. In that sense, the AGI debate is no longer entirely about predicting the future. It is increasingly about determining where we are today.
