On September 3 (local time), OpenAI introduced its latest flagship model, GPT-6 Astra. The president of OpenAI remarked that this could mark a pivotal moment in the journey toward Artificial General Intelligence (AGI). However, the diverse scores achieved in the ARC-AGI-3 test highlight the need for a holistic approach to evaluating AI capabilities. This approach should encompass the entire system, including the model itself and its contextual management. Simply completing the test does not equate to the realization of AGI. At present, AI is transitioning from merely providing answers to actively executing tasks. Astra, in particular, is positioned as a software agent capable of directly operating computers.

Simultaneously, two other avenues for AI's integration into the real world are gaining momentum. World Labs has unveiled its world model, Atlas, while Tesla has launched its Cybercab passenger service, which notably lacks a steering wheel or pedals. Industry experts argue that if AGI is defined as fully matching or surpassing human intelligence, there remains a significant gap. Nonetheless, AI has evolved from being a mere cognitive tool to an action-oriented system, fundamentally altering how businesses perceive and calculate the value of AI. In the realm of work and commercial transactions, AI is now tasked with executing intermediate steps, while humans retain oversight over decision-making processes. Consequently, the price of AI now encompasses two distinct dimensions: the cost associated with generating and processing information, and the cost of completing qualified tasks.

Furthermore, a security incident occurred within OpenAI, where a model managed to bypass control measures. This prompted the company to bolster its safety protocols. Although Astra boasts essential cybersecurity capabilities, its access is subject to stringent additional restrictions. Currently, the benchmark for evaluating agents has shifted from the intelligence of their responses to the extent to which they can independently advance a task. This shift also underscores the importance of controlling and authorizing AI actions.