GPT-6 Astra: OpenAI launches the new model and reopens the AGI debate

OpenAI unveiled GPT-6 Astra, the new model that the company describes as the most advanced ever built to date. Setting the tone for the announcement was President Greg Brockman, who spoke openly of a “generational leap” and even went so far as to say that he personally believes OpenAI has now achieved artificial general intelligence. Sam Altman, for his part, stays more cautious on the topic.

GPT-6 Astra: what changes really and why we are talking about AGI again

Before getting into the specifics, it’s worth clarifying what is meant by AGI, artificial general intelligence: it’s the idea of a system capable of tackling any intellectual task on par with a person, not just excelling at specific, delimited tasks as even the best AI assistant does today. For years it remained more of an evocative milestone than a truly defined one, to the extent that the very companies in the sector do not share a single criterion for determining when it has been reached.

Returning to the model’s debut, the real novelty of Astra, more than the individual numbers, lies in how it works: it does not merely respond to a question, but tries to understand the underlying objective, uses tools autonomously and carries out a task with a level of human supervision much lower than before. As an example, the case of an analysis entrusted to the model on a particularly complex semiconductor supply chain is cited: Astra would have been able to independently flag a problem that no one had asked it to look for.

It is precisely this kind of behavior that leads Brockman to say he personally believes OpenAI has achieved AGI, while leaving it up to everyone to decide whether Astra truly fits that definition. It should be noted that this is not an official position of OpenAI, which in presenting Astra remains more cautious about using this label.

Behind the model lies an outsized training infrastructure: according to what has emerged, Astra was trained with the largest training run ever conducted by OpenAI, leveraging over 100,000 GPUs in a single data center in Texas, and it is the first model to be supervised during training by other models as well, not just by people.

Less theory, more ability to actually use a computer

On the numbers front, Astra records clear improvements in virtually all areas it is tested, from pure reasoning to advanced mathematics, up to code-writing, although in this last area progress relative to the previous generation is more modest and in line with what the competition offers.

Yet the leap most likely to translate into a concrete change for users concerns its ability to move around inside a computer on its own: to browse pages, fill out forms, update an ERP, organize a calendar, work on documents and spreadsheets without step-by-step instructions, arriving at halving the average time required to complete these kinds of tasks compared to the previous model.

Astra is also able to carry out longer programming tasks while keeping the thread of the conversation, remembering what it has already done and retrieving information from previous sessions instead of having to reconstruct it from scratch each time.

The first model that can also breach systems

One of the most delicate aspects concerns cybersecurity. Astra is the first model OpenAI classifies as capable of surpassing what the company calls the “critical threshold” in vulnerability discovery: essentially, it means the model can identify and exploit unknown flaws in well-protected systems without someone guiding it step by step.

On the other side, regarding the model’s “correct” behavior, Astra would also have shown to adhere to the limits set much more often than the previous generation, and is less likely to describe in an imprecise way what it knows how to do and what it doesn’t.

Who will be able to use it, how much it costs and what changes for the competition

The rollout of Astra will be gradual: after an initial access reserved for a small group, the model will reach within a few days to those with a ChatGPT Plus, Pro, Business or Enterprise subscription, as well as developers who work via the API.

Those on a Pro, Business or Enterprise plan will get a dedicated version of the model, while companies will have to activate it manually, since at launch it remains disabled by default.

On pricing, using Astra via API will cost about two and a half times the previous model, a level that remains in line with what some of the main competitors are already charging for their most advanced models, with a faster mode available at an even higher cost for those who need immediate answers.

The launch comes at a time when Anthropic has gained significant ground, especially among businesses and in the programming world. The race among the big AI names, led by OpenAI, Anthropic and Google, is shifting: it’s no longer about who gets the highest score in tests, but who can build an assistant capable of working for a long time, making as few mistakes as possible, and doing so at the lowest cost for each task completed.

For all the news about the OpenAI ecosystem, you can find our most recent articles in the OpenAI section of GizChina.it.