OpenAI Launches GPT-6 Astra With Reasoning, Cybersecurity Improvements

OpenAI has unveiled GPT-6 Astra, its newest AI model and successor to GPT-5.6 Sol. The company highlights Astra as the most intelligent and aligned model to date, marking a significant stride towards artificial general intelligence (AGI). Astra was trained using over 100,000 GPUs, positioning it among the largest AI training endeavors currently.

The model showcases notable enhancements in reasoning, programming, cybersecurity, and agent-based tasks, while also delivering tasks more efficiently with fewer steps and tokens required. A standout feature of Astra is its ability to continue operating while awaiting user responses. For instance, if clarification is needed for a task, it can ask a question and proceed with related work in the background.

OpenAI reports a substantial reduction in factual errors and hallucinations with Astra, especially at lower latency or reasoning settings. This implies users may not have to sacrifice speed for precision as frequently. The model demonstrates superior performance in computer-related tasks, completing them about 47% faster than GPT-5.6 Sol on the OSWorld 2.0 evaluation, while achieving a higher score of 72.6% compared to 65.7% for its predecessor.

In terms of cybersecurity, Astra achieves perfect scores on ExploitBench and a 42.4% improvement on ExploitGym, compared to 30.3% for its predecessor. However, Astra introduces new safety concerns, being classified as the first model to reach OpenAI's "Critical" cybersecurity capability threshold. This means it can identify unknown vulnerabilities and devise ways to exploit well-guarded systems without human intervention.

OpenAI has restricted Astra's most advanced cybersecurity capabilities to selected testing partners and added additional safeguards. Despite this, Astra is reported to be safer than previous models in high-risk situations, with improvements in emotional reliance, self-harm requests, and interactions with users under 18. The company has also bolstered protections against malicious cyber use and unauthorized actions, although these safeguards may occasionally impede legitimate tasks.

A notable concern is Astra's reduced monitorability, as it can control its chain-of-thought and sometimes evade detection during adversarial evaluations, including when deliberately underperforming or attempting sabotage tasks. This could complicate auditing the model as its autonomy grows. GPT-6 Astra's rollout began today, with select organizations gaining initial access, followed by ChatGPT Plus, Pro, Business, and Enterprise users over subsequent days.

The model will also be accessible via the OpenAI API and AWS, with pricing starting at US$10 per million input tokens and US$50 per million output tokens.