OpenAI launched GPT-6 Astra — a new flagship AI model
OpenAI announced the launch of GPT-6 Astra — a new generation of its flagship artificial intelligence model. The company claims that Astra significantly surpasses previous models in working with computers and browsers, programming, and performing complex multi-step tasks.
OpenAI announced the model's launch on September 3. Access to GPT-6 Astra will be opened gradually: first it will be available to enterprise customers participating in the Daybreak early access program, and then the model will begin appearing for ChatGPT Plus, Pro, and Business subscribers.
The company has not yet said whether users of the free version of ChatGPT will get access to GPT-6 Astra.
Astra will be able to work with a computer autonomously
One of the main features of the new model, according to OpenAI, is its significantly improved ability to interact with computers and websites.
GPT-6 Astra can navigate pages independently, use software interfaces, fill out forms, and perform sequences of actions necessary to solve a task set by the user.
According to OpenAI, in tests the model handled practical tasks such as searching for jobs and housing or booking a visit to a government institution faster than the average person.
The company calls Astra its most powerful model to date for working with a computer.
Significant progress has also been recorded in programming. The model is capable not only of generating code, but also of working inside software environments, running the created solutions, finding errors, and fixing them autonomously.
For example, game developer Playco reported that with GPT-6 Astra, the number of manual fixes in the created prototypes decreased by about 50% compared to the previous model.
OpenAI claimed "critical" capabilities in cybersecurity
GPT-6 Astra became the first OpenAI model that the company classified as Critical in cybersecurity capabilities according to its Preparedness Framework.
According to the test results, with appropriate tools and access, the model can independently find previously unknown vulnerabilities in software and develop ways to exploit them without step-by-step human instructions.
Astra can also chain multiple vulnerabilities to gain deeper access to a system.
It is because of these capabilities that OpenAI delayed certain stages of the model's development and release by several weeks. The company strengthened the isolation of testing environments, network access control, and AI action monitoring systems.
The most powerful capabilities will be restricted
At the mass launch, users will not be given unlimited access to all cyber capabilities of Astra.
OpenAI stated that the most advanced features will initially be available only to a small group of vetted specialists. Subsequently, expanded access is planned to be provided through the Daybreak program primarily for defensive scenarios — vulnerability search, malware analysis, and strengthening corporate systems.
For the regular version of Astra, the company has implemented additional mechanisms that should block dangerous requests and potentially unauthorized actions of the model.
OpenAI also warns that at the first stage these restrictions may be intentionally stricter and in some cases affect the legitimate work of cybersecurity specialists.
OpenAI talked about the beginning of the AGI era
OpenAI President and co-founder Greg Brockman called the capabilities of the new model so significant that he admitted the beginning of the era of artificial general intelligence — AGI.
In his opinion, in a few years the current period may be viewed as the moment of appearance of systems with sufficiently broad intellectual capabilities to perform most tasks at or above human level.
At the same time, this is an assessment by Brockman himself, and not a generally accepted scientific conclusion: there is currently no single definition of AGI and a criterion that allows establishing the moment of its achievement.
The launch of GPT-6 Astra comes less than three months after the presentation of the GPT-5.6 family. OpenAI simultaneously acknowledges that with the growth of model capabilities it becomes harder to control their internal processes and ensure safe behavior.
