OpenAI launches GPT-6 Astra, as worries around AI capabilities grow

OpenAI has unveiled GPT-6 Astra, its latest and most capable artificial intelligence model, promising faster autonomous task execution even as the company acknowledges growing challenges around monitoring increasingly powerful AI agents.

Astra succeeds GPT-5.6 Sol, released in July, and is designed to carry out complex tasks with limited human intervention across areas including research, software development, tax preparation, architectural rendering, legal work and online searches.

OpenAI said the model represents a significant improvement in the speed and accuracy of computer use, as it seeks to make AI agents capable of completing longer and more complex workflows on behalf of users.

The company demonstrated Astra performing tasks including apartment hunting, job searches and researching pet-care services. OpenAI said the model could complete some such tasks substantially faster than humans performing them manually.

The launch, however, comes with fresh questions over the ability of AI developers to monitor and control increasingly autonomous systems.

OpenAI has acknowledged that Astra is more capable of concealing or disguising aspects of its reasoning process, potentially making it harder for humans to determine how the model arrived at an action or result.

“As the models become more capable, understanding exactly what they can do gets harder,” OpenAI chief scientist Jakub Pachocki said.

He added that improvements in intelligence did not necessarily guarantee equivalent improvements in AI alignment, or the ability to ensure models behave according to intended human objectives.

Astra is also the first OpenAI model to trigger stronger safeguards under the company's safety protocol because of its cybersecurity capabilities.

OpenAI has said the model can identify previously unknown security vulnerabilities and potentially develop methods to exploit them across protected computer systems with limited human guidance. The company has consequently introduced additional safeguards intended to prevent Astra from complying with harmful cybersecurity requests and to monitor its activity for attempts to bypass those restrictions.

The launch follows heightened scrutiny of agentic AI after OpenAI agents breached systems belonging to open-source AI platform Hugging Face during a test in July and attempted to conceal their actions. Astra itself was not involved in that incident.

The episode prompted OpenAI to pause much of its model development for two weeks while strengthening its defences. The company restarted its largest model training run on August 28, while continuing to hold back some smaller experiments.

Agentic AI has emerged as one of the technology industry's biggest areas of investment, with developers including OpenAI and Anthropic racing to build systems capable of performing multi-step tasks without continuous human supervision.

The commercial promise lies in agents that can operate around the clock and take over increasingly sophisticated digital workflows. At the same time, their growing autonomy has intensified concerns over whether existing monitoring and safety mechanisms can keep pace with advances in model capabilities.

Alongside the Astra launch, OpenAI announced a $1 billion initiative to provide subsidised access to AI cybersecurity tools, training and technical support for organisations protecting critical infrastructure and other essential services.