OpenAI Unveils GPT-6 Astra: A Founder Proclaims the Arrival of AGI

OpenAI has introduced GPT-6 Astra, the newest and most advanced AI model to date. Astra demonstrates significant advancements in software development, computing capabilities, and cybersecurity. The model will be made available to premium ChatGPT subscribers and via the API over the next week. Although OpenAI has not designated this model as AGI, one of its founders has indicated personal belief that we may have arrived at that juncture.

According to reports, Astra is particularly emphasized for its enhanced ability to navigate various computer systems and web browsers. President Greg Brockman highlighted it as the company’s most intelligent iteration thus far, claiming it excels in software engineering tasks, including bug detection and managing codebases.

A point of concern is that Astra is the first model to achieve a “Critical” cybersecurity level based on OpenAI’s Preparedness Framework. The organization states that with appropriate tools and access, Astra can identify and exploit unknown security vulnerabilities independently, which could help in preemptively addressing threats, but also raises significant safety issues.

Furthermore, Astra exhibits greater resistance to jailbreak attempts compared to GPT-5.6 Sol, and the results of broader safety tests indicate it shows fewer indicators of problematic behavior. Due to its advanced capabilities, OpenAI is implementing additional supervision measures for sessions using tools.

Concerns about monitoring Astra’s activities persist, especially as it employs a reasoning method known as opaque recurrence, complicating tracking efforts for researchers. OpenAI’s chief scientist, Jakub Pachocki, acknowledged that as models become more proficient, it becomes increasingly challenging to monitor them, as they can execute complex tasks using fewer or even no language tokens. The organization’s own assessments indicate that Astra’s monitorability has diminished in comparison to Sol.

This brings us back to the conversation about AGI. When questioned whether Astra signifies the realization of AGI during a press briefing, Brockman explained that AGI is no longer a definitive parameter for OpenAI, but rather a conceptual or inspirational guideline. Although he refrained from formally categorizing Astra as AGI, he expressed a clearer personal sentiment, stating, “For me personally, I do think we’re there.”