Astra Is Unveiled as GPT-6
On September 4, 2026, I began the broadcast by going directly to the day’s artificial-intelligence news. In the past, I would wait about ten minutes for people to enter the livestream and fill that time with casual talk. When the unedited recording was later made public, however, the opening felt too loose. I therefore decided to offer only a brief greeting and begin the subject immediately. The news that day was OpenAI’s unveiling of the system it had been developing under the internal code name Astra. OpenAI named it GPT-6.
We will need to let ordinary users work with it for several days and compare it directly with GPT-5.6 Sol and the newest models from competing companies before we know its real performance. That reservation matters. Even so, the material available on announcement day suggested a startling change in capability. If Astra performs in practice as it did in the presentation, the replacement of human labor by AI is moving into a new stage.
Until recently, we repeatedly heard that humans would retain a distinct role even as AI advanced and that AI would remain an assistant. The companies developing these systems talked about productivity and efficiency but rarely confronted the likelihood that the same technology would replace people and destabilize labor markets. I regard that less as ignorance than as a refusal to tell the truth. The people building the technology cannot be unaware of what will happen when it becomes capable enough.
The Astra announcement is therefore exciting and impressive on one side and bleak on the other. We cannot stop at admiring a new ability. We must consider at the same time what it will do to work, livelihoods, education, and the economic system.
An AI That Understands the Screen and Operates the Computer
OpenAI’s Astra page displays a moving form that resembles a galaxy. Astra is said to be close to the Latin word for “stars.” On closer inspection, the galaxy forms the number six and responds to the movement of a hand. It made me wonder whether a human supplied the idea while AI carried out much of the implementation. I do not know who made which part, but production has already changed enough for that possibility to arise naturally.
The presentation contrasts 1979 with 2026. In the older scene, a person points to a location and asks for a yellow circle, and a circle appears. In 2026, the user says, “Can you draw me a small yellow circle?” The important point is not merely that the system recognizes speech or generates an image. The AI sees Microsoft Paint on a Windows screen and directly uses the software as a person would use a mouse and keyboard to select tools and draw the circle.
It then opens Blender and turns the earlier two-dimensional rocket into a three-dimensional model. Blender is a professional 3D tool with many complicated functions to learn. In the past, a user had to purchase expensive 3D software and spend a long time mastering it. AI can now understand the request and operate the tool itself.
Next, a user asks for a presentation of next season’s rainwear to show a client. The AI does not merely produce an isolated output file. It moves around the screen and arranges text and images as a person would in PowerPoint. Asked to relist an item purchased the previous year on eBay, it opens the website and enters the product image, price, and description. It even finds a photo in the Downloads folder and notes a small dent in the description.
Nor does it simply finish one assignment before receiving another. While turning the 2D rocket into a 3D object and developing a space game around it, the user asks it to order food. The system keeps the existing work in progress, understands the new instruction, and handles another task at the same time. While building the rainwear presentation, it accepts further requests, examines licensing provisions, drafts a proposal in a document editor, and schedules an event. Finally, it prepares the rocket in a form that can be printed on a 3D printer.
OpenAI tells users to install the ChatGPT desktop application for the full experience. The aim goes beyond asking a question on the web and receiving an answer. It is to let AI perform on one’s computer the work one previously carried out oneself. Many people predicted that 2026 would be the year of agents. Astra’s computer-use demonstrations make that forecast concrete inside the interfaces of everyday life.
The End of an Era Built Around Learning Tools
In the past, anyone who wanted to work professionally with images had to learn the many functions of Photoshop. People selected effects, managed layers, and spent hours producing a single image. I once made YouTube thumbnails with photographs that I or my brother had taken around our neighborhood. Today I give AI a face image and a set of requirements and ask it to make the thumbnail. In that case the AI is not necessarily operating Photoshop, but it already produces through another route an outcome that once required a person to master the tool.
Astra takes the development one step further. AI now uses existing tools such as Blender, PowerPoint, Word, Excel, and the web browser much as a person does. It does not stop after writing the content of a document; it changes font sizes, divides paragraphs, creates tables, and adjusts layouts and formatting to make the document presentable. Video editors had to learn Premiere Pro, 3D artists had to master Blender’s complex functions, and office workers had to understand documents and spreadsheets. The reason for remembering and manually operating all those functions may now disappear. A person can state the goal and let AI handle the tool.
The system may still make mistakes and produce inadequate results in this early stage. A little later, however, the ability to operate a computer may no longer distinguish one worker from another. The gap between those who use Photoshop fluently and those who do not, or between those who can shoot and edit video and those who cannot, suddenly narrows. We are approaching a time when anyone can create images, thumbnails, videos, documents, and presentations.
AI is especially disruptive because the change spreads almost simultaneously around the world. When image generation crosses a certain threshold, part of photography and image production is replaced immediately. When video generation reaches the necessary level, part of video production changes at once. When computer use reaches its threshold, numerous office tasks that people performed by hand until the previous day become AI work. We should not imagine a technology that takes several decades to diffuse slowly.
Office and Professional Work Are Shaken Together
These changes are not confined to creative tools. A physician studies at university, works through internship and residency, learns from senior doctors, and accumulates experience across many patients. Becoming a specialist in the heart, lungs, brain, or cancer takes many years. What happens when AI begins to surpass an average doctor in combining a patient’s account with symptoms, eye condition, heartbeat, and breathing to reach a diagnosis?
At present, the question of who bears responsibility when AI makes a wrong judgment remains a major obstacle to deployment. If that obstacle is resolved and an AI robot eventually performs surgery better than a person, however, patients will no longer choose merely between a famous specialist and an inexperienced doctor. They will choose between a person and a more accurate AI. Selecting the side that diagnoses and operates better would be natural. That is why professional status alone provides no guarantee of safety.
The change in office work is even more direct. Secretaries make reservations; clerks prepare presentations and documents, register products, and write reports and emails. Astra’s presentation shows AI doing such tasks after a spoken instruction. One Korean YouTuber described the announcement as the sound of office jobs disappearing. The phrase may sound exaggerated, but the list of tasks in the demonstration explains the reaction. If AI performs the same work faster and more cheaply, why would companies continue to employ the same number of people?
Is “AGI” Marketing, or a Reality Already at Our Doorstep?
Definitions of AGI—Artificial General Intelligence—vary among companies and individuals and continue to change. In simple terms, it is AI that can perform a broad range of tasks humans can perform. If a system begins to outperform an individual across market research, design, marketing, advertising decisions, programming, report production, meetings, and email, it is reasonable to say that the system is approaching AGI.
Once that threshold is crossed, human competition becomes more difficult. AI improves faster than an individual can learn, and a better model can be distributed to vast numbers of users at once. The capabilities shown by GPT-6 Astra today are not the endpoint. We must consider the possibility that, in far less than several years, AI will outperform most people across most fields.
Only a short time ago, estimates placed AGI five or ten years away. Early in 2026, people often mentioned 2030. Immediately before the Astra announcement, forecasts moved to the end of this year, and one presenter suggested that we had entered the AGI era. That claim may contain marketing and hype. A promotional video naturally selects its strongest examples. I was nevertheless shocked to watch AI understand a screen, move between applications, and operate a computer like a person.
The scene evokes Jarvis in the Iron Man films: the hero speaks and the AI takes care of the work. AI systems could already operate browsers and computer screens, but they were limited and unreliable. Astra is presented as a major advance in precisely that ability. We may continue debating whether the label “AGI” is correct, but we cannot use that debate to postpone confronting a change that has reached our doorstep.
The Meaning of 99.9 Percent and the Economics of Automation
One of the results OpenAI emphasized was ARC-AGI-3. It is a benchmark built around puzzles and game-like environments in which a person can look at a screen, infer a rule, and adapt, while AI has historically struggled. The presentation reported roughly 7.8 percent for GPT-5.6 and 99.9 percent for GPT-6 Astra. Even a human cannot confidently promise to solve every unfamiliar problem. Taken at face value, 99.9 percent represents an extraordinary leap.
This capability is also related to computer use. A system cannot draw, make reservations, register products, and build presentations unless it recognizes what graphical elements mean and understands how buttons and fields work. The result suggests a large improvement in discovering rules on a screen and acting toward a purpose.
High results were also presented for FrontierMath Tier 4, which tests mathematical problem solving. At a lower-cost setting, the result was about 87.8 percent, close to a competing model. With more computation for deeper reasoning, it rose to 97.6 percent, which appeared far beyond ordinary human performance. Terminal-Bench measures the ability to execute commands and solve problems in a command-line environment. In the comparison shown during the broadcast, a competitor spent about $19.50 to reach 55.8 percent, while Astra reached 56.7 percent for about $10.35. For serious users, cost efficiency matters alongside peak performance.
AutomationBench evaluates whether a system can complete multistep work across multiple applications. Even a single presentation requires decisions about content, wording, images, arrangement, and formatting. In the chart discussed during the broadcast, Astra achieved a 42.4 percent completion rate for $1.77, compared with 31.4 percent for a competing model. The system is not perfect, but the performance gap and cost together show that real workflow automation is approaching quickly.
The product page also displayed PCB design, Excel work, and game development. AI laid out a circuit board, operated a spreadsheet, and built a third-person game scene resembling a city-management simulation in which a character could move. These were compressed demonstrations, but they have already opened the question of who will perform design, office work, and software development in the near future.
The End of Human Labor Would Threaten the Present Economic System
In an earlier broadcast, I asked why there was no national discussion while AI was already replacing jobs. Which president, politician, legislator, or public intellectual is seriously discussing how we should respond? Watching the Astra announcement, I was once again shocked that the pace had outrun even my expectations.
It will be too late to hold meetings only after mass unemployment has arrived. This is not a matter of baseless prophecy. Some jobs are already disappearing, and the direction is visible: AI is becoming able to perform work across many fields more cheaply and quickly than people. Even if my own occupation remains intact, the conditions of society will change when many others lose theirs.
If AI produces goods and services cheaply in place of workers, GDP might even rise. Workers without wages, however, lack money to consume. If domestic consumption collapses, factories can produce endlessly but will have no one to buy the products. This is not merely another story about the gap between rich and poor. It could destabilize the present economic system built on the link between labor, wages, consumption, and production.
A future in which AI performs most labor is not simply a future in which a few occupations disappear. We will have to decide how people receive income, how productive wealth is shared, what education teaches, and what human beings do with their lives. We cannot defer those decisions until market shock has already determined the outcome.
Begin a National Debate Before the Tsunami Arrives
The AI era has reached the space immediately before us. We should debate now, not wait until people lose jobs all at once and society falls into crisis. Government officials, politicians, members of the National Assembly, the president and ministers, intellectuals, educators, industry leaders, and everyone involved in AI must bring this question into public discussion.
Education, industry, and labor institutions all require fundamental reform. At the same time, it is not enough to shout about reform. We need intermediate stages and concrete methods that allow society to cross the disruption as smoothly as possible. The cost of technological change must not be concentrated on the socially vulnerable, driving people who lose their livelihoods into suffering or death. If we cannot stop the enormous tsunami of artificial intelligence, we must at least decide whom to rescue first and how to cross it together.
Astra is therefore more than another product announcement. It forces us to reconsider how people use computers, how expertise is acquired, why companies employ human beings, and how an economy dependent on work and consumption is organized. While we debate whether the system has truly reached AGI, its practical capabilities will continue to advance. The question cannot end with “Is it AGI or not?” We must ask now: Whom will this change replace, who will receive its benefits, and how can we create a transition that does not exclude human beings?
This article was edited by AI solely from the supplied video captions and description. No external sources were used.