Assessments are mounting that OpenAI's newly unveiled artificial intelligence model, GPT-6 Astra, has ushered in an "Astra moment" — four years after ChatGPT arrived in late 2022. As capabilities that previously required multi-agent systems became readily available, expectations for the mass adoption of AI agents surfaced alongside warnings that, in the words of some observers, "as early as three years from now, the job landscape centered on routine office work will face structural dismantling." Some also predict that demand across AI infrastructure, including memory semiconductors, will explode as computing architecture and inference methods are reshaped, moving beyond the earlier approach of simply expanding parameter counts and extending inference time.
Astra, which Nvidia Chief Executive Jensen Huang on the 7th declared had reached artificial general intelligence, is delivering striking results in perceived performance that go beyond developer benchmark scores. Its rollout to general subscribers paying $20 a month, in addition to the $200-a-month premium tier, is spreading user experience rapidly.
On developer platforms, reviews continue to pour in: "With just a few prompts to Astra, I built a game with graphics on par with the latest 3D titles to near-completion in 45 minutes," and "It fixed system errors and code conflicts in a single run that existing coding-specialized models couldn't resolve after repeated correction instructions." High productivity in actual work settings is also evident. On social media, it is easy to find examples of Astra completing image and video editing and multi-file Excel and presentation tasks almost like a human — work that other agents such as Claude previously failed to deliver clear results on.
Astra is also posting record-setting results in game play, used to gauge an AI's problem-solving ability. It finished the role-playing game Pokemon FireRed in 18 hours and 12 minutes using only screen analysis, without outside walkthrough information. Its predecessor GPT-5.6 Sol took 96 hours to reach the ending, while GPT-5.5 failed to clear the game even after more than 200 hours. Human players who do not know the walkthrough need 20 to 25 hours. The AI taught itself how to progress through a complex RPG from scratch and finished faster than a typical person.
According to the developer community GitHub, Astra was also the only model evaluated to score a perfect 450 points on the 2026 College Scholastic Ability Test, answering questions in Korean, English, mathematics, Korean history and four elective subjects.
Whereas AI previously stopped at producing a single answer to a query, Astra runs code itself and verifies errors through internal loop-based inference until it meets its goal, delivering a finished product. Simply logging into the ChatGPT app or website now allows users to run AI agents that experts previously could deploy only by building complex multi-agent setups. The technical basis is an approach called recurrent depth. Rather than enlarging the model, it deepens inference and logical capability through repeated recomputation. After improvements in pretraining-centered model performance hit a ceiling, the development paradigm shifted to test-time computing, which emphasizes the inference process; recurrent depth marks another evolution, adding repetition to that inference.
Because Astra achieves high performance by repeating computation until it reaches its objective, the computing capacity required is also expected to rise sharply. Even though the price per token — the unit of AI computation — has fallen through model efficiency gains, the total tokens consumed per task has surged, pulling in infrastructure demand more forcefully. Bandwidth and transmission speed are expected to gain further value in particular, given the need to continuously loop real-time data input and output. That is why AI accelerators, high-bandwidth memory (HBM) bandwidth and ultra-high-speed network performance are drawing attention. Hwang Soo-wook, an analyst at Meritz Securities, said: "Astra is significant in that it lowered the barrier to entry for AI agents, as a model that keeps going until it produces a result when given a goal." He added that the hardware demand base will broaden as required AI computation expands.
The outlook is also brightened by the fact that OpenAI's next model is already in training with a far larger computational budget. OpenAI Chief Executive Sam Altman, introducing the development process for the next model after Astra's release, disclosed that training had been paused at one point to examine problems of controlling intelligence. Huang likewise said that more than 100,000 Blackwell NVL72 units went into training Astra and that 400,000 graphics processing units will run for the next model — four times the infrastructure deployed for Astra.
Semiconductor market forecasts are improving sharply with Astra's arrival, and the memory supply cliff is coming into clearer view. KB Securities estimated that finished memory inventories at Samsung Electronics (005930) and SK hynix (000660) have plunged to less than 10 days as of the third quarter. Some observers even suggest that output is shipping as soon as it is produced and that available volumes for sale could be entirely exhausted next year. Goldman Sachs also forecast that global memory demand is being met at only about 60%, that the supply-demand imbalance will deepen in 2027 and that shortages will persist through 2028.
