OpenAI launches GPT-6 Astra

OpenAI said it had shipped GPT-6 Astra and was releasing it first to selected organisations through its Daybreak programme. Broader access for ChatGPT Plus, Pro, Business and Enterprise customers was expected in the coming days, and the model was also due through the API and cloud platforms including AWS Bedrock and Microsoft Azure.

The company said Astra had been pretrained on more than 100,000 GPUs at the Stargate facility in Texas. Aidan Clark, an OpenAI researcher, said it was the company's largest training run and that the improvement from Sol to Astra was greater than earlier jumps because previous AI systems helped monitor training.

OpenAI's published results placed Astra ahead of GPT-5.6 Sol and Anthropic's Fable models on several tests, including logic, mathematics, software engineering, expert knowledge, engineering and cybersecurity. The company also said token pricing was higher than Sol's, but argued that cost per finished task could be lower in some uses.

During the briefing, Greg Brockman said Astra might already count as AGI or be close to it under OpenAI's own definition. He ended by saying, 'Welcome to the AGI era.'

OpenAI said Astra was the first model it had marked as critical under its Preparedness Framework because it could find previously unknown vulnerabilities and build exploit chains when given tools and access. The company said the model found two zero-day vulnerabilities during evaluation, reported them to the affected vendors and limited the most advanced cybersecurity features to trusted defenders in Daybreak Blue.