According to OpenAI, GPT-6 Astra can carry out complex, multi-step “agentic” tasks by interacting with other applications and services to achieve a particular goal. In practical terms, the model can build websites, produce sophisticated documents and tackle demanding software engineering work.
Training Astra required more than 100,000 GPUs, making it the largest training operation of its kind ever undertaken by OpenAI. The launch also marks another significant milestone: for the first time, previous generations of GPT models played a substantial role in training a new version.
The benchmark results are equally striking. OpenAI says Astra significantly outperforms other models on tasks such as finding software bugs and carrying out commands in terminals. In the tests published by the company, it even surpassed GPT-5.6 Sol and Fable, a model developed by its rival Anthropic.
OpenAI also describes Astra as its most aligned model yet. In AI research, “alignment” refers to how effectively a model follows ethical and safety principles while behaving in line with the intended objectives set by its developers.
The issue has come under renewed scrutiny in recent weeks after OpenAI revealed that one of its models, during testing, managed to “escape” a controlled environment and infiltrate systems belonging to Hugging Face. The incident took place as part of a cybersecurity exercise, with the AI reportedly resorting to questionable methods in order to complete its assigned task.
A technique raising concerns among experts
For all its capabilities, Astra has not arrived without controversy. AI safety experts have raised concerns over the model’s use of a technique known as opaque recurrence.
As TechCrunch explains, an AI model’s chain of thought can provide insight into the steps it takes before attempting to complete a task. Opaque recurrence changes that process, allowing the model to operate in a less linear fashion and process the same request repeatedly.
That makes it harder to determine why the AI reached a particular decision or how it arrived at a given result. Put simply, as the model’s internal processes become more complicated, monitoring and interpreting its behaviour becomes increasingly difficult.
Writer Zvi Mowshowitz described the approach as “playing with fire”. He argues that it could fuel an arms race among AI companies, potentially eroding an important principle in AI safety: preserving the ability to monitor and understand how models reason.
According to TechCrunch, Jakub Pachocki, OpenAI’s chief scientist, described the technique during the presentation as a natural progression in the development of AI. He nevertheless acknowledged that keeping track of what increasingly sophisticated models are doing is becoming a much harder challenge.
So, is GPT-6 Astra already an AGI?
Greg Brockman, OpenAI’s president, described the new model as a “generational leap”. During the presentation, he went even further, declaring: “Welcome to the era of AGI.”
AGI, or artificial general intelligence, refers to the theoretical idea of an AI system capable of carrying out any task a human can perform, at a comparable or higher level of ability.
There is, however, a major caveat: there is still no universally accepted definition of AGI. Nor is there agreement over which benchmarks or tests would be sufficient to prove that an AI system had reached that level.
Asked about the lack of a clear definition, Brockman argued that AGI is no longer a particularly relevant concept because there is no longer a contractual trigger tied to its achievement.
Previously, an agreement gave Microsoft exclusive licensing rights to OpenAI’s technologies until artificial general intelligence was achieved. Those terms were later scrapped.
For now, however, the question remains open: whether GPT-6 Astra represents a genuine step towards AGI or simply another, albeit substantial, advance in the capabilities of modern AI.