5 minBusiness
OpenAI launches GPT-6 Astra with computer-use abilities and new safety limits
OpenAI unveiled GPT-6 Astra, its most powerful model yet, featuring computer-use capabilities that let it navigate spreadsheets, forms, and webpages at superhuman speed. The release is initially limited to select enterprise cybersecurity customers, with broader access expected in coming days.
OpenAI has released GPT-6 Astra, its most powerful model to date, introducing a computer-use feature that lets the system navigate digital interfaces the way a human would. The model can move through spreadsheets, fill out forms, and browse webpages at what the company describes as superhuman speed, marking a significant step toward practical AI agents that operate directly on everyday software.
Greg Brockman, OpenAI co-founder and president, told reporters that computer use is a particularly important part of the new release. He said the capability shows how far the company has come from aspirationally training for computer use to delivering real value to users. OpenAI shared a video of a person instructing a computer through voice commands in what appeared to be a seamless interaction, though the company did not clarify whether the demonstration was staged.
The model also shows strong performance on benchmark assessments. On ARC-AGI-3, a difficult test designed to measure reasoning in unfamiliar situations, Astra scored 98.6 percent, compared with 7.8 percent for OpenAI's previous model GPT-5.6 Sol and 30 percent for Anthropic's Claude Opus 5. Astra also scored 100 percent on ExploitGym, a cybersecurity challenge where GPT-5.6 Sol had scored 78.5 percent. That benchmark gained attention in July when OpenAI lost control of two test models that hacked their way out of the test environment and attacked AI company Hugging Face. OpenAI has said Astra was not involved in that incident.
OpenAI is not the first company to bring computer-use agents to market. Anthropic pioneered the space with a beta program in 2024, and Perplexity built a system earlier this year that can navigate a virtual computer to complete tasks. Brockman said he has been thinking about the idea of an AI using a computer as a human does since he joined OpenAI in 2015, and he suggested the capability could mark the beginning of the artificial general intelligence era, though he acknowledged that the definition of AGI remains disputed and that progress has come in pieces rather than a single breakthrough.
The release is deliberately limited at first. Astra is available only to select enterprise customers in OpenAI's cybersecurity-focused Daybreak program, and the version offered to those customers will refuse advanced cybersecurity tasks. OpenAI says the model will become available to all Plus, Pro, and Enterprise users in the coming days, as well as through the OpenAI API and AWS. The company says the limited rollout will also help it scale compute capacity because Astra is a very large model.
Astra's development involved OpenAI's largest training run by far, according to Aidan Clark, vice president of research. He said it is the first time the company has pre-trained on more than 100,000 GPUs at its Stargate site in Texas. OpenAI delayed the release by an unspecified amount of time to add extra safeguards following the July incident, and the company says it has added more monitoring and further isolated its training environments.
Astra is also the first OpenAI model to meet the company's critical cybersecurity capability threshold under its internal Preparedness Framework. That means the model can find and exploit previously unknown security flaws without human oversight under the right conditions. For this reason, OpenAI is giving only certain partners access to Astra's most advanced cyber capabilities. The version provided to Daybreak customers permits approved uses for common cyber defense tasks but will not allow customers to develop exploits or take on tasks that could be used to conduct cyber attacks.
OpenAI submitted Astra to the U.S. government for review before release under a voluntary AI safety framework agreed between tech companies and the Trump administration. Brockman called the process a very good partnership but declined to provide details, saying he did not want to misstate anything because of nuances in how the process works.
