Nvidia chief executive Jensen Huang has declared that “AGI has arrived” following the release of OpenAI’s GPT-6 Astra, placing one of the technology industry’s most influential executives behind an unusually strong claim about the progress of artificial intelligence.
In a post on X on September 6, Huang congratulated OpenAI’s team and highlighted the speed at which the company had moved from the original ChatGPT to its o1 reasoning model and now Astra within four years.
Huang said GPT-6 Astra was trained using more than 100,000 Nvidia Grace Blackwell GPUs connected through NVLink72 systems. He added that another 400,000 GPUs were “coming online next,” although he did not specify where the additional processors would be deployed, who would operate them or when they would become available.
His statement followed OpenAI’s September 3 introduction of GPT-6 Astra, which the company describes as its most intelligent and aligned model. Astra is designed for computer use, online research, software engineering, cybersecurity, scientific work and complex professional tasks.
Astra Expands Beyond Answering Questions
GPT-6 Astra is built to carry out longer digital workflows instead of responding only with text.
OpenAI says the model can complete activities such as filling online forms, updating customer records, conducting web research, installing and testing software and organising information across workplace applications. It can also create documents, spreadsheets, presentations, websites and data analyses based on instructions and existing templates.
The company reported that Astra scored 72.6% on its version of OSWorld 2.0, a benchmark that evaluates how effectively AI systems operate computers. GPT-5.6 Sol scored 65.7% under the same setup.
Astra also completed the evaluated computer-use tasks in approximately 40 minutes on average, compared with roughly 75 minutes for Sol. OpenAI said this represented a 47% reduction in the time required per task, although the results come from controlled simulations and may not reflect every real workplace environment.
For software development, Astra scored 57.9% on Terminal-Bench 4.0, compared with 37.3% for GPT-5.6 Sol. It also reached 97.6% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3.
These benchmark results indicate substantial progress on selected evaluations, but they do not independently establish artificial general intelligence. Scores can vary according to model settings, tools, prompts and the testing environment.
OpenAI itself has not formally certified Astra as AGI in its launch announcement. The company describes the model as a new generation of intelligence and its most capable broadly deployed system, but “AGI has arrived” remains Huang’s assessment rather than an agreed scientific conclusion.
What Artificial General Intelligence Means
Artificial general intelligence generally refers to an AI system capable of learning, reasoning and performing successfully across a wide range of intellectual tasks rather than specialising in one narrow area.
However, there is no universally accepted definition or single test for determining when AGI has been achieved. Some definitions focus on matching humans across cognitive tasks, while others require an AI system to perform most economically valuable work at or above human ability.
A model can therefore produce high benchmark scores and still have weaknesses involving factual accuracy, unfamiliar situations, judgment, memory or real-world reliability.
Huang’s statement is commercially significant because Nvidia supplies much of the computing infrastructure used to develop and operate advanced AI models. Every increase in model size, training activity and user demand can create additional requirements for graphics processors, networking equipment and data-centre systems.
His reference to another 400,000 GPUs coming online suggests that spending on frontier AI infrastructure is continuing after Astra’s training. However, Huang’s brief post did not explain whether those processors are intended exclusively for OpenAI or will be distributed across several Nvidia customers.
Cybersecurity Capabilities Bring Added Restrictions
Astra is OpenAI’s first model to reach the “Critical” cybersecurity capability level under the company’s Preparedness Framework.
OpenAI says that, with appropriate tools and access, the model can identify previously unknown security weaknesses and develop methods for exploiting protected computer systems without requiring a person to guide every step.
That capability could help security professionals identify vulnerabilities before attackers use them. It also creates a risk that malicious users could attempt to employ the model for sophisticated cyberattacks.
OpenAI has introduced stricter safeguards, including risk-based access controls, monitoring of tool-using sessions and systems designed to stop potentially unauthorised actions. These protections may occasionally pause or block legitimate defensive-security work while the company evaluates a request.
The company also reported that Astra is more resistant to jailbreak attempts and more likely to stay within an authorised task than GPT-5.6 Sol. At the same time, OpenAI acknowledged that the model can be harder to monitor in adversarial tests because it has greater control over the reasoning displayed to internal monitoring systems.
Rollout Begins Across ChatGPT and Cloud Platforms
GPT-6 Astra initially became available to a limited group of organisations. OpenAI plans to expand access to ChatGPT Plus, Pro, Business and Enterprise users, along with customers using its application-programming interface.
The model will also be offered through Microsoft Azure and Amazon Web Services’ Bedrock platform. Pro, Business and Enterprise customers are expected to receive access to the higher-capability GPT-6 Astra Pro version.
Standard API pricing has been set at $10 per million input tokens and $50 per million output tokens. A faster processing mode will cost twice the standard rate.
The launch strengthens OpenAI’s position in the competition with Google, Anthropic and other model developers. It also reinforces Nvidia’s role as the hardware supplier behind the industry’s largest training projects.
Whether Astra represents AGI will remain disputed. What is clearer is that it combines stronger computer control, professional-task performance and cybersecurity capability with a computing requirement large enough for Nvidia’s chief executive to place the hardware scale at the centre of his congratulatory message.
