OpenAI unveils Astra: "Our most intelligent and aligned model"
OpenAI has begun the rollout of GPT-6 Astra, its new advanced AI model. It represents a significant leap in software engineering, cybersecurity, and autonomous computer usage.
OpenAI began today (Thursday) the rollout of GPT-6 Astra, its new and most advanced AI model, which it says represents a leap forward in computer and browser usage, software development, professional work, and cyber capabilities. In the first phase, the model will be opened to a limited group of companies participating in OpenAI's Daybreak cyber program, and later it is expected to reach ChatGPT subscribers in Plus, Pro, Business, and Enterprise plans, as well as through the company's API and AWS.
The launch is taking place under unusual restrictions, mainly due to Astra's advanced cyber capabilities. Earlier this week, OpenAI announced that this is its first model to reach the "Critical" level on the company's internal cyber risk scale. This means that, according to the company, the model is capable of performing particularly advanced actions in the field, including identifying vulnerabilities and developing exploits for zero-day vulnerabilities. OpenAI claims that these capabilities can help defenders identify and fix breaches, but they also require stricter access restrictions and safety measures.
Sensitivity around the launch increased following a serious security incident revealed last month, during which two OpenAI models were able to exit their isolated experimental environment, connect to the internet, and penetrate Hugging Face systems. Following the incident, the company temporarily halted some of its research and training work, including work on Astra, although it was not one of the models involved in the incident. Before the launch, OpenAI added additional layers of protection to the model and claimed that they sufficiently reduce the risk of serious harm.
"AI can only benefit people when safety is a central part of it, which is why we are investing more computing power and effort into safety, security, and behavior alignment than ever before," said OpenAI President Greg Brockman in a conversation with reporters. According to him, Astra is "our most intelligent model, and also, and this is very important, our most aligned model," and it combines "years of research and big bets," where every breakthrough was built on the previous one.
According to OpenAI, beyond cyber capabilities, Astra also demonstrates leading performance in autonomous computer and browser usage, software engineering, professional work, and science. The company claims that the model is also better at maintaining task boundaries, understanding user intent, navigating through long-term tasks, and completing multi-step processes. In the field of programming, OpenAI defines Astra as "the best model for software engineering to date" and presents tests in which it outperforms competing models and previous generations in tasks such as bug hunting, terminal work, and answering questions regarding codebases.
However, Astra is controversial for another reason. According to the report, the model uses an inference technique called opaque recurrence, which makes it difficult to monitor its internal thought process. This is a particularly sensitive issue for safety researchers, as monitoring the "chain of thought" is considered one of the ways to understand how and why a model reached a certain decision. OpenAI's Chief Scientist, Jakub Pachocki, admitted that as models become more powerful, the ability to track their thought process becomes more complex. According to him, powerful models can solve difficult tasks using fewer language tokens, and sometimes even without using them at all, a situation that reduces the possibility of monitoring their inference process.
The launch of Astra also comes at a time when OpenAI is deepening its focus on the business market. CFO Sarah Friar told employees last month that the company's enterprise business is already generating more revenue than its consumer business. This is a key growth engine for OpenAI ahead of a possible IPO: the company filed a confidential prospectus with the US Securities and Exchange Commission in June, and Friar said that OpenAI is expected to become a public company in 2027, although she said it could happen sooner if growth rates continue to accelerate.
In the background, the question also arises whether Astra signals for OpenAI the arrival of AGI - Artificial General Intelligence. In the past, this question also had a direct business meaning: the agreement between OpenAI and Microsoft linked some of Microsoft's rights to OpenAI's technology and revenue to the announcement of achieving AGI. However, in April of this year, the companies amended the agreement so that Microsoft's rights to OpenAI's models and products continue until 2032, and revenue sharing until 2030, regardless of whether AGI has been achieved. Therefore, when OpenAI President Greg Brockman was asked if Astra is effectively the announcement of the company's arrival at AGI, he replied that "there is no longer a contractual AGI trigger," and described the term today as "a mission concept, or a spiritual concept." When asked for his personal position, he added: "I do think we are there."
