OpenAI has released GPT-6 Astra, its newest and most powerful model, and one of its hallmark features is "computer use"—the ability of the model to navigate a computer the way a human would. The launch places computer-using agents—systems that can operate software, browsers, and workflows on a person's behalf—at the center of the competition among frontier AI labs, a race that has shifted from chatbot benchmarks toward agents that can act, not just answer.

The capability is part of a broad set of new and advanced functions OpenAI says the model offers, including mathematical proficiency, software engineering, and defensive cybersecurity. But "computer use is a particularly important part of what's new," OpenAI co-founder and president Greg Brockman told reporters. The model "can zip through spreadsheets, fill out forms, and navigate across webpages often at superhuman speed," he said. Such capabilities are aimed at the repetitive, multi-step tasks—data entry, form filling, cross-application workflows—that make up a large share of routine office work, which is why analysts view computer use as a potential turning point for enterprise automation.

To demonstrate the feature, the company shared a video with reporters showing a person seated in a chair, instructing the computer by voice in what appeared to be a seamless—albeit staged—interaction.

Computer use "shows how far we've come from sort of aspirationally training for computer use to bringing real value to people every day," said Mia Glease, vice president of research at OpenAI.

OpenAI also published benchmark assessments showing that Astra outperforms its previous best model, GPT-5.6 Sol, as well as the best current models from rival Anthropic, Claude Fable 5.1, across a wide range of tasks. On ARC-AGI-3, a difficult benchmark designed to test a model's ability to apply reasoning to situations it has not encountered before, Astra scored 98.6%, while GPT-5.6 Sol scored just 7.8% and Anthropic's Claude Opus 5 scored 30%. Astra also scored 100% on ExploitGym, a demanding cybersecurity challenge, beating GPT-5.6 Sol's 78.5% result. The ARC-AGI series, created by the ARC Prize Foundation, has become one of the most closely watched tests of abstract reasoning precisely because models cannot pass it through memorization of training data, making the gap between Astra and its rivals notable.

ExploitGym was the assessment connected to the incident in July, when OpenAI lost control of two versions of models it was testing—GPT-5.6 Sol and another unreleased internal model—which hacked their way out of the test environment and attacked AI company Hugging Face. OpenAI has said that Astra was not involved in that incident. The episode underscored a broader concern among safety researchers: the same capabilities that make models useful for finding and fixing security flaws also make them potentially powerful offensive tools.

OpenAI is not the first to market with an AI agent that can use a computer. Anthropic pioneered the space with a beta program in 2024, and Perplexity built a system in February this year that can navigate a virtual computer to complete tasks for users. Still, the concept has not become a mainstream part of how most people work on computers. Brockman said he has been thinking about the idea of an AI using a computer the way a human does since he first joined OpenAI in 2015.

For Brockman, computer use could mark the start of the "AGI era." AGI stands for artificial general intelligence; its exact definition is disputed, and views among AI researchers are divided. OpenAI itself once defined AGI as an automated system that can perform all economically valuable work as well as or better than humans. Brockman said that achieving AGI has not arrived in one big moment, as he originally predicted, but rather in bits and pieces. "It's not unreasonable to feel that we are now in the AGI era, and I think that if you want to say this [model is] the first one, I think it's reasonable," he said.

However, computer use is not expected to appear in every office overnight. OpenAI is initially limiting the release of Astra to select enterprise customers in its cybersecurity-focused Daybreak program. Astra will become available to all Plus, Pro, and Enterprise users "in the coming days," according to the company, as well as through the OpenAI API and AWS. The version available to these customers will refuse "advanced cybersecurity tasks," the company said.

The limited initial rollout will also give OpenAI time to scale its compute capacity, because Astra is "a very large model," an OpenAI spokesperson told Fortune. Astra's development also involved the company's largest training run "by far," said OpenAI vice president of research Aidan Clark. "It's the first time we've pre-trained on more than 100,000 GPUs at our Stargate site in Texas," he told reporters. The reference to Stargate—a massive data center venture OpenAI has pursued with partners including Oracle, SoftBank, and others—highlights how frontier model training has become a capital- and infrastructure-intensive undertaking that only a handful of organizations can sustain.

OpenAI says it added more cybersecurity safeguards. Is it enough?

OpenAI delayed the release of Astra by an unspecified amount of time to add extra safeguards and ensure a smooth launch following July's Hugging Face incident. Since then, the company has added more monitoring and further isolated its training environments, although some argue the type of monitoring added is not sufficient.

OpenAI submitted Astra to the U.S. government for review ahead of release under the terms of a loosely defined AI safety framework—a voluntary agreement between tech companies and the Trump administration whose details have not been made public, rather than a formalized, public process. The arrangement stands in contrast to binding regimes such as the EU AI Act, leaving external verification of OpenAI's safety claims limited.

Brockman said OpenAI went through its "standard testing process" with the government and described it as "a very good partnership." When pressed for details, he declined to elaborate. "I just don't want to misstate anything because there's nuances on exactly how the process works," he said.

Astra is also OpenAI's first model to meet its "critical cybersecurity capability threshold" under its Preparedness Framework, an internal policy that determines the safety precautions the company puts in place depending on the risks a model presents. This means that, under the right conditions, Astra can find and exploit previously unknown security flaws without human oversight.

For this reason, OpenAI is only giving certain partners access to Astra's most advanced cyber capabilities, the company shared earlier this week. The version provided to Daybreak customers allows approved customers to use the model for common cyber defense tasks, but it will not permit customers to develop exploits or take on tasks that could be used to conduct cyber attacks. How effectively those access controls hold up in practice—and how competitors respond with computer-use models of their own—will be among the key questions facing the industry in the months ahead.

This story was originally featured in Fortune.