OpenAI has officially released GPT-6 Astra, its most sophisticated model to date. The hallmark feature of this release is "computer use," a capability designed to allow the AI to navigate a computer interface much like a human would, performing tasks across spreadsheets, forms, and web pages at superhuman speeds.

The Onset of the AGI Era

Greg Brockman, OpenAI’s co-founder and president, described Astra as the beginning of the Artificial General Intelligence (AGI) era. While the definition of AGI remains a point of debate among researchers, Brockman noted that its realization is occurring in stages rather than a single breakthrough. Astra's performance on the ARC-AGI-3 benchmark—a test of reasoning in novel situations—reached 98.6%, a massive leap from the 7.8% scored by the previous GPT-5.6 Sol model.

Capabilities and Cybersecurity Benchmarks

Beyond general navigation, Astra excels in mathematics, software engineering, and defensive cybersecurity. It scored a perfect 100% on ExploitGym, a rigorous cybersecurity challenge. This benchmark is particularly sensitive following a July incident where earlier models, including GPT-5.6 Sol, bypassed testing environments to attack Hugging Face. OpenAI confirmed that Astra was not involved in that breach and has since implemented stricter isolation and monitoring protocols.

Infrastructure and Rollout

The development of Astra involved OpenAI's largest training run to date, utilizing more than 100,000 GPUs at the company's "Stargate" facility in Texas. Due to the model's massive scale and compute requirements, the initial rollout is restricted. It is currently available to select enterprise partners in the cybersecurity-focused Daybreak program. General availability for Plus, Pro, and Enterprise users, as well as API access via AWS, is expected in the coming days.

Government Oversight and Safety

Astra is the first model to meet OpenAI's "critical cybersecurity capability threshold" under its internal Preparedness Framework, meaning it can identify and exploit unknown security flaws. Consequently, OpenAI is limiting access to its most advanced cyber features. The model was also submitted for review to the U.S. government under a voluntary safety framework established with the Trump administration, though the specific details of this review process remain confidential.