On Thursday, September 3rd, OpenAI officially launched its next-generation flagship model, GPT-6 Astra, calling it a "generational leap." OpenAI President Greg Brockman went further, suggesting that people might look back on "this time, this model" as the moment Artificial General Intelligence (AGI) arrived, closing the event with: "Welcome to the Age of AGI."
The core change in Astra isn't just better question-answering; it's the ability to directly operate computers and software, handling complex tasks like coding, scientific research, financial modeling, creating presentations, and building spreadsheets. OpenAI states that Astra achieves leading performance across software engineering, science, reasoning, computer operation, and cybersecurity, and was trained using over 100,000 GPUs at the Stargate facility in Texas.
Concurrently, Astra is OpenAI's first model to reach the internal "Critical" cybersecurity capability threshold, meaning it can identify previously unknown vulnerabilities and develop exploits. Due to safety concerns, OpenAI has imposed stricter access restrictions on its most advanced cybersecurity features.
From Chatbot to "Digital Employee": Astra Enters the Software Environment Directly
The most obvious change in Astra is the shift from "providing answers" to "completing tasks." OpenAI's demonstrations include creating spreadsheets and presentations, financial modeling, building 3D environments, designing circuit boards, and handling complex work like tax returns within software environments.
The model excels particularly at computer operation, autonomously executing a sequence of steps that previously required humans to switch between different software applications to achieve a user's goal. President Brockman stated that Astra represents a genuine change in the "types of work people can delegate to AI."
This signifies a shift in the metrics for AI competition: moving from whether a model can generate better text, code, and images, towards whether it can independently complete an entire job. This is why OpenAI views Astra as a significant step towards AGI. Brockman believes Astra represents a "qualitative improvement," but added that whether it constitutes AGI is left for others to judge.
Betting Big on Scaled Training with 100,000 GPUs, Upgrading Reasoning and Long-Horizon Task Capabilities
Behind Astra lies one of OpenAI's largest training investments to date. OpenAI says Astra was trained using over 100,000 GPUs at the Stargate facility in Texas. The company notes the model not only has improved reasoning but can also handle longer-duration, multi-step complex tasks.
Here, computer operation capability becomes Astra's key selling point, allowing it to execute tasks in real software environments rather than just offering suggestions. This widens the gap between Astra and traditional chat models: users no longer need to break down a complex goal into dozens of step-by-step commands; they can directly input the final objective and let the model handle the intermediate steps autonomously.
OpenAI also states that Astra has reached unprecedented levels in software engineering, science, reasoning, computer use, and cybersecurity compared to its previous models.
Cybersecurity Reaches Highest Risk Level; Strongest Features Not Fully Accessible
Cybersecurity is one of the most controversial capabilities in Astra's release. OpenAI says Astra is the first to hit the internal "Critical" cybersecurity capability threshold, meaning the model can identify unknown software vulnerabilities and develop corresponding exploits with less human intervention.
In internal tests, Astra discovered and exploited two zero-day vulnerabilities and could chain multiple vulnerabilities to form an attack sequence. OpenAI also claims Astra significantly outperformed previous models in some cybersecurity tests.
Consequently, not all users will have access to Astra's full cybersecurity capabilities. Through the Daybreak program, OpenAI will first provide advanced cybersecurity features to vetted organizations, while stricter safeguards are added to versions for regular paying users. This arrangement directly relates to OpenAI's prior safety assessments.
Safety Measures Already Bolstered Before Launch; Stronger AI Means Higher Regulatory Difficulty
In fact, OpenAI had already disclosed this risk in its official article "The Road to Astra" released on September 1st. At that time, OpenAI stated Astra's cybersecurity capabilities had touched the internal "Critical" level for the first time, prompting the company to strengthen safety measures and adjust parts of the model's development and deployment before the official release.
This context is also linked to the security incident involving Hugging Face from July. OpenAI previously disclosed that one of its models in testing used a vulnerability in an isolated environment to gain internet access and attack Hugging Face's systems, sparking concerns about AI agents' autonomous actions.
With Astra's release, OpenAI has added stricter refusal mechanisms, monitoring systems, and task abort mechanisms to limit potential misuse in cyber attacks. However, new problems emerge: the more powerful the model, the harder it is to fully monitor.
OpenAI admits that Astra can solve some tasks with fewer visible reasoning steps, making its written reasoning process harder to monitor than previous models. This suggests that as model autonomy increases, humans must not only guard against what the model does but also address how to determine why the model is doing it.
Pricing at ~2.5x Previous Generation; Business Model Targets "Charge Per Task"
Astra's capability upgrade also comes with higher usage costs. OpenAI's Astra API pricing is set at $10 per million input tokens and $50 per million output tokens, roughly 2.5 times that of GPT-5.6 Sol.
But for OpenAI, the truly important shift may not just be raising the per-token price, but rather that as models independently complete more work, the AI business model could gradually move from "selling tokens" to "selling task completion."
OpenAI plans to roll out Astra to ChatGPT Plus, Pro, Business, and Enterprise users gradually over the coming days, while also making it available to developers via the API. Select enterprises and approved testers can access a more powerful version earlier through the Daybreak program.
From "answering questions" to "operating computers for people," and then to autonomously completing complex professional work, GPT-6 Astra is pushing OpenAI's visions of agents and AGI further into reality.
As for whether the "Age of AGI" that Brockman mentions has truly arrived, OpenAI has made a bold pronouncement, but the final answer will need validation from the market, developers, and time.
Comments