"The Smart Newcomer Never Stops Working... This 'Rookie' Is Intimidating [Click e-Industry]"

The Computing Race Ignited by GPT-6
"AI Now Works Independently for Hours"

Beyond Basic Q&A: AI Handles Real-World Tasks
Computing for Long-Term Reasoning Grows in Importance

The launch of OpenAI's new artificial intelligence (AI) model, 'GPT-6 Astra,' is expected to greatly increase the amount of time AI can independently perform work, while also significantly altering the competitive landscape of the AI industry, according to analysts in the securities industry.

A newly hired "rookie" who recently joined the company is rapidly expanding their scope of work. There is a growing expectation that artificial intelligence (AI), which operates independently without human intervention, will establish itself as a new workforce. The photo is an image representing a new employee. Photo by Getty Images

On September 7, Lee Youngjin, a researcher at Samsung Securities, stated, "The most significant change with GPT-6 Astra is the expansion of the time horizon in which AI can autonomously carry out tasks to completion without human intervention." He added, "The standard for competition in the AI market will go beyond simply possessing a model, to encompass how much real work can actually be processed."

Expansion of Autonomous Execution Time...Execution Environment Design Determines Performance

With the launch of OpenAI's new artificial intelligence (AI) model 'GPT-6 Astra,' it is anticipated that the time AI independently performs tasks without human intervention will significantly increase. Getty Images

Astra has enhanced its ability to use computers and tools for extended periods, perform complex multi-step tasks, and employ multiple agents simultaneously. Beyond simply achieving correct answers for singular problems, the ability to autonomously complete work over several hours is now the factor driving substantial differences in productivity.

Researcher Lee noted, "Astra is the first model to achieve a 'Critical' level in OpenAI’s safety evaluation framework for cybersecurity." He explained, "This advancement goes beyond simply knowing about vulnerabilities; Astra can now actually navigate systems, discover new vulnerabilities, and solve problems based on its improved capabilities."

He went on to emphasize the importance of execution environments, referencing ARC-AGI-3 evaluation results. Lee said, "Even with the same model, there was about a 37 percentage point difference in performance depending on state management and context retention." He added, "In future AI competition, not only the model’s inherent capabilities but also how memory, tools, and the execution environment are designed to enable stable long-term operation will become important factors."

AI Accelerates AI Research...Continued Rise in Demand for Computing Infrastructure

As AI is now being actively deployed in research work to develop next-generation AI, a clear acceleration in research is emerging. At OpenAI’s research organization, agent work equivalent to about 3.1 human workdays is being executed per researcher workday. Each researcher operates several agents in parallel, delegating coding, debugging, experiment execution, and result analysis to them.

Lee explained, "A flywheel effect is emerging, where increased research throughput and productivity are leading to further improvements in AI performance." He added, "As human time ceases to be a bottleneck, computing resources will become increasingly important as the main constraint on the pace of research."

Even if improved model efficiency reduces the tokens and costs required per task, overall inference and computing demand could still surge explosively as the range of work that AI can handle expands and more parallel agents are utilized.

AP Yonhap News

Shift to 'Cost per Task' in Value Assessment

The software industry, too, is facing structural changes. As agents now perform entire business processes while working across multiple software platforms, companies are challenged to evolve their products beyond simple functionality to become the key execution environments for these agents.

Lee said, "The criteria for measuring AI value have already shifted from comparing prices per token to evaluating costs per task." He added, "In the future, key metrics for assessing the economics of AI will include work success rates, task completion costs, and labor hours that can be saved—rather than just price per token."

This content was produced with the assistance of AI translation services.