OpenAI has released the GPT-6 Astra model, marking a pivotal shift from an Answer Engine to an AI Agent. Astra can now operate computers, browse the web, use software, write and execute code, and complete complex workflows, achieving a 72.6% success rate on the OSWorld 2.0 benchmark—surpassing GPT-5’s 65.7%. This signifies AI’s evolution from “answering better” to “successfully getting things done,” transforming it from a mere question-answering tool into a digital workforce that can be “hired.” Token demand will shift from simple user growth to a multiplicative effect driven by three variables: the number of agents, the number of tasks per agent, and token consumption per task. AI is now forming a flywheel effect of “Human + AI → Train → Better AI,” where improvements in model capability itself become the engine driving increased compute demand.Article author and source: Minority Viewpoint
Almost everyone would agree with the following statement:
In the age of AI, what will defeat you is the person or company that adopts AI earlier and uses it more effectively.
But after OpenAI launched its own GPT-6 Astra model, the above statement is completely inaccurate—or entirely wrong. The truly correct statement now is:
In the AI era, what will outpace you is no longer those who simply use AI better, but those who proactively acquire, deploy, and truly harness AI agents.
Just a reminder: the "you" mentioned above can also refer to a company.
Do you remember? Last month, Astra made a brief appearance as an internal secret model at OpenAI, delivering top-tier mathematical research—on par with Fields Medal-level work—on ten mathematics and theoretical computer science problems. This time, OpenAI’s official release of Astra marks its biggest breakthrough: it’s not just that Astra has become “smarter” with enhanced mathematical and scientific capabilities—it’s that Astra has transformed AI from an Answer Engine into a true AI Agent.
Therefore, with this release, OpenAI no longer limits itself to discussing how much Astra's benchmark scores have improved; instead, it clearly defines the boundaries of Astra's capabilities from the outset:
Astra can operate computers, browse the web, use software, write and run code, complete complex workflows, and perform long-range tasks in science, cybersecurity, and professional settings.
This sentence summarizes Astra’s true breakthrough: AI is now truly evolving from “human → AI → answer” to “human → AI agent → task → execution → result.” Therefore, GPT-6 Astra means:
AI is no longer just a tool for answering questions—it is beginning to become a digital workforce that can be "hired" to work endlessly without rest.
Astra's enhanced capabilities in long-range missions are the most critical.
In OSWorld 2.0—a benchmark proposed by teams from CMU, Stanford, and others that evaluates AI agents' ability to complete complex tasks such as office work, multi-file software engineering debugging, and system-level cybersecurity configuration, followed by a baseline comparison with human experts—Astra achieved 72.6%, surpassing GPT-5.6 Sol's 65.7%. Meanwhile, the simulated time to complete tasks decreased from approximately 75 minutes to about 40 minutes.
This means Astra is undergoing a highly decisive change:
Moved from “answering better” to “successfully getting things done.”
This would completely transform how we understand "AI penetration rate" and how we calculate total token demand. In the past, when estimating the size of the AI market, we naturally asked: How many people are using ChatGPT? How many corporate employees are using Copilot? What is the penetration rate of AI among programmers, customer service representatives, white-collar workers, and scientists?
Then, estimate the size of the AI market using “number of users × ARPU.” This logic was certainly valid in the Answer Engine era. But in the AI Agent era, problems arise:
- One person using AI does not mean only one AI workload; a single person can create multiple Agent tasks; an enterprise can deploy hundreds or even thousands of Agents simultaneously. An Agent can run continuously for hours, even 24 hours a day.
Therefore, the following inequality holds:
1 person ≠ 1 AI user ≠ 1 unit of AI workload.
A materials scientist can configure dozens of AI agents at once to carry out research tasks across different directions. Therefore, the key variable to focus on becomes:
- How many AI agents are currently running worldwide?
- How many tasks does each running AI Agent perform per day? How many tokens does it consume?
The product of these two numbers will be the true demand function for AI.
Whether or not you use AI, whether you observe your friends using AI, or whether your company uses AI—these factors are irrelevant to the adoption rate of AI and the total demand for tokens. If you choose not to use AI, someone else will inevitably use AI agents to replace your advantages or your business, and ultimately replace you and your company.
Therefore, let me specifically add this:
Today, at the very least, all mid- to large-sized companies should prioritize developing their own AI strategy at the board level, as there are too many smart people eyeing your profitable businesses.
Therefore, the definition of AI penetration has shifted from "Per Human" to "Per Agent." It was once thought to be this way:
AI penetration rate = Number of people using AI / Potential population
After Astra, the formula for AI agent penetration is as follows:
Agent penetration rate = Tasks taken over by AI agents / Tasks eligible for agent execution
The denominators of these two formulas naturally differ by a huge order of magnitude, since agents no longer need to correspond one-to-one with humans—an individual may simultaneously own multiple personal AI agents. As for the number of “tasks that can be executed by agents,” I believe OpenAI’s recent introduction page for Astra provides a hint.
Above is a star map from OpenAI's recent Astra launch page; if humanity's insignificance in the universe highlights our profound ignorance of the cosmos, then every exploration of that ignorance becomes a potential task for AI agents—leading to the conclusion that the number of AI agents could be limitless.
So, "how many people worldwide are starting to use AI" and "how many digital workforces are currently operating" are two entirely different orders of magnitude.
The total demand for the token has greater elasticity.
The overall demand for tokens will exhibit significantly greater elasticity than the growth in AI Agent user numbers, because demand in the Agent era does not simply rise with the number of users—it also increases with the amount of AI workloads each Agent handles.
Therefore, the total demand for the token can be roughly expressed as:
Total token demand ≈ Number of AI agents × Tasks per agent × Tokens consumed per task
And all three of these variables may grow rapidly at the same time.
First, an increase in the number of agents. In the past, one person might have used only one ChatGPT; in the future, a single person may simultaneously run personal assistant agents, coding agents, research agents, finance agents, shopping agents, and more. Companies may scale from dozens or hundreds of agents to thousands or even tens of thousands.
Second, the tasks each agent performs are increasing in complexity. In the past, a request like “Summarize this article” might have required only a few thousand tokens; in the future, it could be “Research this industry, read 500 documents, compare 50 companies, build a model, detect anomalies, generate a report, and then provide investment recommendations based on the report.” This is where Astra’s significance lies: the more complex tasks a model can accomplish, the more confident humans will be in entrusting such tasks to agents.
Third, the Agent will shift from "occasional invocation" to "continuous operation." This is the most important change. Traditional ChatGPT operates only when a user asks a question, provides an answer, and then ends. In contrast, an AI Agent is
Receive target → Plan → Search → Execute → Verify → Modify → Execute again → Complete.
It is a continuously running computational chain that constantly generates API calls. The more complex the task, the longer the reasoning, the larger the context, the more tool calls, the greater the output, and the higher the token consumption.
Therefore, AI agents will not only increase the number of AI users but also amplify the token demand behind each AI user.
Therefore, it is entirely possible for the number of AI agents to increase by only 20%, while the total demand for tokens grows by 300%.
The RSI will give Astra a stronger flywheel effect.
Astra is OpenAI's largest training run to date, with publicly available information indicating that its training scale has surpassed 100,000 GPUs. More importantly, AI is increasingly involved in the model training process itself.
It used to be "Human → Train → AI," and is gradually becoming:
Human + AI → Train → Better AI
This is already the basic mode of RSI, where the model can participate in generating training data, identifying issues, conducting evaluations, providing feedback, and even assisting with certain engineering and research tasks during the training process.
Thus, a very compelling flywheel effect chain will be formed:
Stronger models enable more capable AI agents, which can accomplish a greater number of real-world tasks, generating higher-quality task data and feedback that, in turn, help train even stronger models—further enhancing agent capabilities and enabling them to take on more tasks...
This means AI development will not follow the "version iteration" model of traditional software, but rather resemble a self-accelerating production system. As a result, demand for tokens will experience a multiplicative effect driven by multiple simultaneous factors, and the stronger the model, the more pronounced this multiplier effect becomes.
So, we will definitely see:
The improvement of model capabilities itself becomes the engine driving growth in AI computing demand, leading to billions, tens of billions, and ever more AI agent digital workers, which in turn unleashes an endless surge in demand for tokens.
This is the significance of the GPT-6 Astra release.