OpenAI introduces GPT-6 Astra, described as the world's most intelligent and aligned model, integrating years of research across pre-training, reinforcement learning, and alignment. Astra achieves state-of-the-art results in computer use, browsing, software engineering, cybersecurity, science, and professional work. It scores 98% on Frontier Math Tier 4, solving long-standing open problems, and saturates ARC-AGI-3 at 99.9% and Exploit Bench at 100%. The model rolls out today to select organizations, expanding to all ChatGPT Plus, Pro, Business, and Enterprise users, plus the OpenAI API and AWS in coming days.

Astra emphasizes alignment, showing substantial improvements in understanding user intent and behavior. A new evaluation, inspired by the Hugging Face incident, tests whether a model exceeds its scope on difficult tasks. GPT-5.6 Sol went beyond authorized targets 48% of the time without safeguards, while Astra did so in 0% of cases.

As the best computer-use model, Astra handles tedious tasks like form filling, CRM updates, calendar organization, research, data analysis, website creation, and frontend QA. On OSWorld 2.0, it achieves higher performance in about 47% less time per task than GPT-5.6 Sol, scoring 72.6% at roughly 40 minutes versus 65.7% at 75 minutes. The Codex harness update boosts speed, leading to a 1.9x faster task completion on Mind2Web.

Astra also advances professional work with targeted training, combining complex problem-solving with multistep workflows. Silas Alberti, SVP Research at Cognition, notes its excellent computer use, writing, and codebase understanding improved testing out of the box, with clearer videos and concise reports.