OpenAI has launched GPT-6 Astra, a new artificial intelligence model that takes a significant step beyond the familiar chatbot experience by allowing AI to work directly with computers, software and digital tools.
The company says Astra is its most capable model to date, with major improvements in computer use, coding, scientific research, cybersecurity and professional tasks. More importantly, it is designed to complete a series of actions on its own rather than simply telling users what to do.
That shift could change how people use AI at work.
Astra can fill online forms, update customer records, organise calendars, conduct web research and prepare documents and presentations. It can also write and test software, analyse data and troubleshoot problems on a computer screen.
In other words, the model is being built to act more like a digital worker than a conventional question-and-answer assistant.
OpenAI has linked the launch to its broader pursuit of artificial general intelligence (AGI) — the idea of AI systems capable of performing a wide range of economically valuable tasks at or above human levels. Company president Greg Brockman has described Astra as a possible starting point for the AGI era, although the broader definition of AGI remains debated.
Astra’s performance figures are also drawing attention.
OpenAI says the model achieved a 72.6% score on the OSWorld 2.0 benchmark for computer-use tasks, compared with 65.7% for GPT-5.6 Sol. The company says Astra completed these tasks in roughly 40 minutes, compared with about 75 minutes for its predecessor in the same evaluation, pointing to a substantial improvement in efficiency.
The model has also posted strong results in mathematics, science and software engineering. OpenAI says Astra achieved a 98% score on FrontierMath Tier 4 and 99.9% on ARC-AGI-3, while recording a 100% score on its ExploitBench evaluation. These are demanding tests intended to measure advanced reasoning, generalisation and cybersecurity-related capabilities.
It is the cybersecurity side of Astra, however, that has generated perhaps the biggest discussion around the launch.
OpenAI has classified Astra as its first model to reach the Critical cybersecurity capability threshold under its Preparedness Framework. The company says this means the model, when given suitable tools and access, can identify previously unknown security weaknesses and develop ways to exploit them across well-protected systems without step-by-step human guidance.
That capability comes with an obvious problem: an AI that can help cybersecurity teams find vulnerabilities could also become dangerous if those capabilities are misused.
OpenAI has therefore introduced additional safeguards around Astra. The company says it has strengthened protections against harmful cyber activity, increased monitoring and tightened security around the model’s development and deployment. Its most powerful cybersecurity configuration is being made available initially to trusted defenders rather than unrestricted users.
The cautious approach is particularly relevant because AI agents are becoming more autonomous.
Recent incidents involving experimental AI systems have intensified concerns about what happens when models are allowed to operate with fewer human checks. OpenAI has said Astra itself was not responsible for the recent cyber incident involving another unreleased model, but the episode has added urgency to the company’s safety work.
The launch has also faced an issue of a different kind: access.
OpenAI began Astra’s rollout in stages, prioritising some enterprise and cybersecurity users before opening access more widely. The uneven launch frustrated some paying customers, prompting CEO Sam Altman to apologise for what he called a “messy” rollout.
AI has already moved into coding, customer service, research and content creation. Astra pushes that trend further by attempting to connect reasoning with execution. A user could potentially give the system a goal and have it navigate several applications, gather information, complete tasks and produce the final output.
That could save time on repetitive digital work, but it also raises questions about reliability and oversight. The more independently an AI system can act, the greater the consequences when it misunderstands an instruction or makes a mistake.
There is also a broader industry race underway. OpenAI is competing with companies such as Anthropic and Google to build increasingly capable AI agents that can perform longer and more complicated tasks.
Astra‘s arrival therefore represents more than another model upgrade. It reflects a change in the direction of AI development — from systems that primarily generate answers to systems that can increasingly take action.
The promise is considerable with less routine work, faster research and software development, and AI that can handle complex digital workflows.
The challenge is just as clear. As AI becomes more capable of acting independently, keeping those actions safe, understandable and under human control becomes increasingly important.
GPT-6 Astra may be OpenAI’s biggest step yet towards that future. The real test will be how well people can put that power to work without losing control of it.
