Foreign media believe that after OpenAI launched GPT-6 Astra, the competitive logic of AI has shifted. The focus is no longer just on “who can use AI better,” but on “who deploys and leverages AI agents first.” The article views Astra as a milestone marking the transition from answering questions to executing tasks.
From response to execution
Astra is described as being capable of operating computers, browsing the web, using software, writing and running code, and completing longer workflows. The article suggests this means AI is shifting from “human asks, model answers” to “human delegates tasks to agents, agents complete them.”
In the author’s view, Astra’s key significance lies not just in its enhanced capabilities, but in its growing resemblance to a digital workforce that can be “hired.” It is no longer merely generating answers—it is beginning to take on ongoing work tasks.
Penetration rate is no longer calculated per capita.
The article uses test results from OSWorld 2.0 to show that Astra performs better on long-range tasks and significantly reduces completion time. The author concludes that the calculation of AI penetration must change and can no longer be estimated simply by the number of people using AI.
The more important variables have become: how many agents are running, how many tasks each agent performs, and how many tokens each task consumes. The article suggests that an individual can deploy multiple agents simultaneously, and businesses may run hundreds or even thousands of agents at once.
Token demand will be amplified.
The author further points out that token demand in the Agent era will not only grow with the number of users but also rise in tandem with the complexity and duration of tasks. A continuously running Agent will repeatedly perform planning, searching, executing, verifying, and correcting, leading to more calls and higher token consumption.
The article also notes that the model's participation in training itself creates a flywheel: stronger models lead to more capable agents, which generate more high-quality data, which in turn drives further model improvement. The authors believe this will push demand for compute and tokens into a stronger phase of self-reinforcement.