OpenAI presents its most advanced model and Greg Brockman openly talks about the “era of AGI.” But what does general artificial intelligence really mean, and why does GPT-6 Astra mark an important milestone even though we cannot yet say that AGI has been achieved?
Add us to your preferred sources.
GPT-6 Astra has arrived. OpenAI has introduced its new flagship model, describing it as the smartest and most aligned the company has ever created. But this launch is not just about higher benchmarks, better answers, or a new generation of ChatGPT. Around Astra, a much broader discussion is opening up: have we entered the era of AGI, general artificial intelligence?
The words destined to inevitably spark debate were spoken by Greg Brockman, OpenAI’s president and co-founder, who during the model’s presentation openly spoke of the beginning of the “AGI era”. An important declaration, especially considering that OpenAI was founded in 2015 precisely with the goal of developing Artificial General Intelligence. But talking about the AGI era doesn’t necessarily mean that an AI equivalent to human intelligence has already been achieved.
To understand why GPT-6 Astra represents an important milestone, we must start with this question: what is AGI?
What is AGI, general artificial intelligence
AGI is the acronym for Artificial General Intelligence, in Italian intelligenza artificiale generale. The concept describes an AI capable of tackling a wide variety of problems, learning, adapting to new situations and transferring knowledge from one task to another, instead of being designed to perform only specific tasks. It is an important distinction. For many years AI systems have been extremely effective within very narrow fields: recognizing images, translating texts, playing chess, classifying information, or making predictions.
The emergence of large language models has progressively widened these boundaries. A system like ChatGPT can already write, translate, code, analyze documents, interpret images, browse the web and tackle very different problems. But the idea of AGI goes beyond this.
A general AI system should be able not only to answer a question but to understand a goal, decide what steps are necessary to achieve it, use different tools, handle unexpected events and carry on the work with a high level of autonomy. And it is precisely on this ground that GPT-6 Astra becomes particularly interesting.
GPT-6 Astra does not just want to answer: it wants to work
OpenAI presents Astra as its most capable model for complex professional tasks and for so-called computer use. The model can use browsers and apps, fill out online forms, update data in a CRM, organize a calendar, conduct web searches, work on documents and create websites. It can also tackle tasks composed of many steps and produce documents, spreadsheets, and presentations as a final result. It’s a seemingly subtle but substantial difference.
With early chatbots, the user would ask a question and the AI would return an answer. With agentic systems, the user can instead indicate a desired outcome and let the AI execute an increasing part of the process needed to achieve it. In other words, we are moving from AI used as an assistant to AI used as an operating agent. This is probably the most concrete meaning of the expression “AGI era,” even more than a single benchmark.
Why OpenAI speaks of the AGI era
AGI does not have a universally recognized threshold within the scientific community. There is no official test to establish that a system has achieved general artificial intelligence once it surpasses a certain percentage. OpenAI itself has used different definitions over time. Historically, it described AGI as a highly autonomous system capable of outperforming humans in most economically relevant activities, while more recently the concept has been more generally associated with systems able to reach or surpass human capabilities in numerous domains.
For this reason, Brockman’s statement should be read carefully. Speaking of “AGI era” does not necessarily mean declaring that GPT-6 Astra is definitively AGI. It may instead indicate entry into a phase where the distinction between specialized AI and a system with much more general capabilities is progressively less clear.
The ARC-AGI-3 benchmark and Astra’s result
One element that fueled the debate is the result achieved by GPT-6 Astra on ARC-AGI-3, a benchmark designed to evaluate systems’ ability to handle new situations, understand rules, and find a solution without having encountered exactly the same problem before. According to the results presented at launch, Astra can reach 98.6% using the environment optimized by OpenAI. In a standardized configuration used to more directly compare different models, the result is around 63%.
The difference between the two values is important because it shows how much the performance of modern artificial intelligence systems now depends not only on the model but also on the infrastructure around it: memory, available tools, context management, and software used to orchestrate operations. And it is also the reason why it would be incorrect to automatically transform a very high result into a certificate of achieving AGI. A benchmark measures certain abilities under defined conditions. The real world is much more complex, ambiguous, and unpredictable.
Is GPT-6 Astra therefore a true AGI?
We cannot say that yet.There is currently no sufficient scientific consensus to establish that GPT-6 Astra represents general artificial intelligence, nor is surpassing one or more benchmarks enough to prove it. The most interesting datum is probably another: some of the characteristics we have used for years to imagine AGI are progressively appearing in systems actually available.
The ability to tackle problems belonging to different domains, use tools, adapt to new tasks, plan complex activities, and greater operational autonomy are all features that reduce the gap between traditional specialized models and the concept of general intelligence. The real shift in era might therefore not coincide with the day when someone can officially announce that a machine has become as intelligent as a human being. It could be the moment when we stop using artificial intelligence just to get answers and start delegating entire processes to it.
The leap into professional work
It is precisely in work that the consequences of this evolution could become visible more quickly. OpenAI has specifically trained GPT-6 Astra to tackle complex professional tasks. The model combines reasoning abilities with computer usage and can execute workflows composed of multiple stages. This means AI does not necessarily have to stop at producing a text or an analysis. It can, at least in supported contexts, gather information, work on different applications, modify documents, and reach the production of the requested result.
For companies and professionals this shift is significant because it progressively changes how work processes will be designed. It will not only be about asking which activities can be accelerated by artificial intelligence, but which processes can be partially entrusted to autonomous systems and which must continue to include human control, responsibility, and decision-making..
More autonomy also means more risks
The very characteristics that make Astra more powerful, however, also make the issue of security even more important. GPT-6 Astra is the first OpenAI model to have reached the Critical level for cybersecurity capabilities as defined by the company’s Preparedness Framework. According to OpenAI, with appropriate tools and access, Astra can identify previously unknown vulnerabilities and develop new ways to exploit them even on strongly protected systems, without a person having to guide every single step.
For this reason, the company states that it has strengthened the protections applied to the model before deployment. This is another element that helps understand why the debate about AGI does not concern only the power of artificial intelligence. The more a system becomes capable of acting autonomously, the more important issues become such as control, supervision, security, responsibility, and the ability to understand why it took a certain action.
When will GPT-6 Astra be available
The launch of GPT-6 Astra began with a gradual rollout. OpenAI initially indicates access for selected organizations through the Trusted Access program. The extension to ChatGPT Plus, Pro, Business, and Enterprise and API access are expected progressively in the days following the launch. Developers will have at their disposal a model with a context window of 1.05 million tokens and up to 128,000 output tokens. The model also supports tools such as Web search, file search, computer use, image generation, code interpreter, shell, and MCP. The API price indicated by OpenAI is 10 dollars per million input tokens and 50 dollars per million output tokens.
From GPT-5 to GPT-6 Astra the AI’s role changes most notably
The evolution becomes even more evident by observing how quickly ChatGPT has changed. Already with the arrival of GPT-5 and subsequent OpenAI model generations the focus shifted from a simple chatbot to systems capable of reasoning and using tools. GPT-6 Astra takes this transformation one step further. The point is no longer just to establish whether an answer is better than that produced by the previous generation. The question becomes how long artificial intelligence can work, how many tools it can use, how many intermediate decisions it can manage, and how reliably it can achieve a complex goal.
The AGI era may already have started without us realizing it
Establishing today whether GPT-6 Astra is or is not a true general artificial intelligence risks being less useful than observing the transformation already underway. Artificial intelligence is rapidly moving from a system that generates content to a system that can act within the digital tools we use daily. This does not automatically make Astra as intelligent as a human being and does not eliminate limits, errors, or the need for supervision. But it changes the relationship between people and machines.
If by “AGI era” we necessarily mean the precise moment when an artificial intelligence equivalent to humans is born, we do not yet have sufficient elements to claim we have reached it. If instead we mean the beginning of a phase in which increasingly general systems can reason, use tools, and autonomously carry out significant portions of human work, then Brockman’s phrase appears much less futuristic.
And perhaps this is the real news behind GPT-6 Astra: the discussion about AGI no longer concerns only something that might happen in the future. It is starting to concern tools that are entering our computers and work processes today.
Sources: OpenAI, official GPT-6 Astra documentation and GPT-6 Astra System Card.
Pubblicato in Artificial Intelligence
Be the first to comment