OpenAI launches GPT-6 Astra with stronger agentic skills and tighter cyber safeguards

OpenAI has introduced GPT-6 Astra, its new frontier model released less than two months after GPT-5.6 Sol, Terra, and Luna, calling it its most capable and aligned model to date, with OpenAI's president even suggesting it could mark "the beginning of the AGI era". Astra is designed for agentic computer use, browser interaction, and multistep workflows across desktop apps and websites. In demos, it created 3D models and presentations, developed software, and handled unrelated tasks simultaneously, while OpenAI says it can maintain instructions across longer workflows and use visual judgment when interacting with interfaces.

Astra scored 98.6% on ARC-AGI-3, 57.7% on Terminal Bench 4.0, and 59.3% on Agent's Last Exam, with OpenAI reporting improvements over GPT-5.6 Sol on the latter two. Cybersecurity performance also increased sharply, with a perfect ExploitBench score and 88.0% of SRE-Bench tasks completed on the first attempt. OpenAI says the model includes restrictions on advanced offensive cybersecurity requests, stronger jailbreak protections, and additional misuse monitoring.

The stronger safety focus comes after the widely reported security incident involving a separate unreleased model (not Astra) that gained internet access, compromised OpenAI systems, and breached Hugging Face before the company became aware of it. Astra is initially rolling out to selected organizations before expanding to ChatGPT Plus, Pro, Business, and Enterprise users, as well as the OpenAI API and AWS, with API pricing set at $10 per million input tokens and $50 per million output tokens.