OpenAI Launches GPT-6 Astra, With Greg Brockman Calling It Possible AGI Arrival

OpenAI released GPT-6 Astra on Thursday, with company president Greg Brockman personally declaring he believes the model constitutes artificial general intelligence — while stopping short of making it an official company position, instead inviting users to judge for themselves whether it clears that historic bar.

Astra was built on OpenAI's largest-ever training run, requiring more than 100,000 GPUs at its Stargate facility in Texas. That scale of compute signals a significant escalation in the AI arms race, and the infrastructure investment alone marks a new threshold for what frontier model development looks like.

Unlike its predecessors, Astra is engineered to operate directly inside software and execute tasks autonomously — not merely suggest what a human operator should do next. That shift toward embedded, action-taking AI agents brings the technology meaningfully closer to handling complex professional workflows end-to-end, a capability that has long been the practical benchmark for AGI in enterprise contexts.

Not everything is being released to the public at once. Astra's most powerful cybersecurity capabilities are being restricted to a small circle of trusted testers — a signal that OpenAI is wary of how certain features could be misused. Safety concerns are also shadowing the launch more broadly, with significant debate in AI circles over whether Astra was trained in ways that could make it and future models more difficult for humans to monitor and interpret.

The release lands during an unusually competitive week, with Anthropic, Meta, and Google all announcing major model updates nearly simultaneously. Whether Astra's AGI framing holds up against real-world performance — and whether it can execute advanced tasks without critical errors — remains an open and consequential question.