OpenAI has launched GPT-6 Astra, a new artificial intelligence model that the company itself considers a significant advancement. Greg Brockman, the company's president, stated that this system has the potential to signal the beginning of Artificial General Intelligence (AGI).
Astra was designed to perform complex professional tasks autonomously. However, this functionality raises an important question related to the evolution of AI agents: how to monitor systems that acquire increasing capacity to act independently?
Brockman expressed his belief that OpenAI may have achieved AGI with Astra, although he left the final determination on this classification to the users themselves. When questioned about the possibility, the executive stated: 'It might be with this model,' summarizing his view by saying: 'Welcome to the age of AGI' during a conversation with journalists.
The training process also reached an unprecedented scale for OpenAI, utilizing over 100 thousand GPUs at the Stargate complex in Texas. Furthermore, for the first time, other models participated significantly in supervising this training.
Unlike previous models that only guided the user, Astra was developed to operate directly within software. Among the capabilities demonstrated or mentioned by OpenAI are tasks such as researching food options and booking a tennis court simultaneously while performing other activities.
Astra's aptitude in cybersecurity is a particularly delicate aspect of the launch. OpenAI categorized the model as 'critical' level in its risk preparation structure. According to the company, the system has the ability to identify unknown flaws and create methods to exploit them in protected systems, even without receiving detailed instructions from an individual.
This concern led OpenAI to pause training for two weeks and implement additional tests and security layers before releasing it. The tests also revealed a distinct challenge: Astra became more difficult to monitor in evaluations designed to verify whether the model could escape human surveillance.
Despite this, OpenAI acknowledges that the model still has difficulties in concealing the reasoning used in complex tasks, but the reduction in monitorability was considered a serious problem by the organization.
Amelia Glaese, OpenAI's Vice President of Research, emphasized the need for trust, stating: 'When models can do more things autonomously, we need to be able to trust them more.' Initial access to Astra will be granted to organizations participating in the Daybreak Access program. It is expected that in the coming days, the model will be available to Plus, Pro, Business, and Enterprise customers, as well as API developers; non-paying users do not yet have a scheduled access date.
The introduction of GPT-6 Astra poses a question to OpenAI that transcends the mere performance of the new model: the greater the autonomy of artificial intelligence, the more urgent the search for reliable methods to track its actions and decisions becomes.