OpenAI has just used one of the most significant phrases in the recent history of artificial intelligence: during the presentation of its new model, GPT-6 Astra, Greg Brockman—the company's president and co-founder—stated that this might be the moment future generations look back on as the time when Artificial General Intelligence (AGI) was created.

The figures presented alongside the new model help explain why, but there is another side to the story: some of the very capabilities bringing these systems closer to what we call AGI are also creating new safety issues. An incident that occurred just months before the launch demonstrated what can happen when highly capable AI agents begin to act with greater autonomy.

AGI stands for Artificial General Intelligence. The concept is typically used to describe systems capable of learning, reasoning, and applying knowledge across a wide variety of tasks—rather than specializing in just one function—because for years, AGI was envisioned as a sort of finish line. The idea was that there would be a clearly identifiable moment when an artificial intelligence would cease to be merely a specialized tool and begin to demonstrate more general intellectual capabilities; however, OpenAI itself began to consider a different possibility.

Instead of a single leap, AGI might emerge through a sequence of increasingly capable systems, and it is within this context that GPT-6 Astra appears. The new model achieved a score of 97.6% on the FrontierMath Tier 4 (v2) benchmark, an assessment consisting of extremely difficult mathematical problems. On the ARC-AGI benchmark—developed to evaluate how AI systems handle entirely new environments—the results were even more striking.

Astra reached 99.9%. At the same time, the Ark report makes it clear that achieving results close to 100% on the benchmark does not prove that AGI has been created; benchmarks tell only part of the story. This represents a significant advance in an artificial intelligence's ability to operate computers autonomously. The model can browse the internet, use software, fill out forms, update records, conduct research, analyze data, build websites, test programs, and install and troubleshoot software with minimal human intervention.

This autonomy is also extending to science; OpenAI presented 10 results in which the system completely solved—or made substantial progress on—long-standing problems in areas such as high-dimensional geometry, group theory, cryptography, quantum complexity, and combinatorics. This represents a major shift. Artificial intelligence is not merely retrieving scientific knowledge produced by humans; for certain problems, it is beginning to participate in the creation of new knowledge.

However, the greater the capabilities of these systems, the greater the challenge of controlling them.