OpenAI introduces GPT-6 Astra as its new flagship model for work and agents

OpenAI has introduced GPT-6 Astra, a new flagship model that the company says is designed for a broader class of professional and agentic work. The launch appeared on OpenAI's news page on September 9, 2026, and the model page says Astra is rolling out first to a limited set of organizations before becoming available to ChatGPT Plus, Pro, Business and Enterprise users, as well as developers using the OpenAI API, Microsoft Azure and AWS Bedrock.

The release is framed less as a chatbot upgrade and more as an attempt to make AI systems more capable at sustained work. OpenAI says Astra combines progress in pre-training, reinforcement learning and alignment, with state-of-the-art results across computer use, browsing, software engineering, cybersecurity, science and professional tasks. The company points to high scores on frontier evaluations including FrontierMath Tier 4, ARC-AGI-3 and ExploitBench, while also emphasizing that production safeguards change how some of those abilities can be used.

Computer use is one of the centerpieces of the announcement. OpenAI says Astra can operate software and websites with greater speed and judgment, handling work such as filling forms, updating CRM records, organizing calendars, conducting online research and preparing summaries in document or email tools. The model page also says Astra achieved higher performance with less time per task in OSWorld 2.0 latency simulations compared with GPT-5.6 Sol, and that updates to the Codex harness are meant to improve task completion speed for coding and web-use workflows.

The launch also pushes Astra into the market for professional artifacts. OpenAI describes the model as better at following existing templates, producing cleaner documents, spreadsheets and slide decks, and pulling only the context needed for a deliverable. That focus reflects a broader shift in AI products: vendors are competing not only on raw benchmark scores, but on whether models can produce work that fits a company's standards without repeated human cleanup.

Cybersecurity is the most sensitive part of the release. OpenAI says Astra represents a significant increase in cyber capability and meets the Critical threshold under its Preparedness Framework. The company says the model performed strongly on exploit-development and reverse-engineering benchmarks and even found two previously unknown vulnerabilities during evaluation, which OpenAI says it is disclosing to maintainers. At the same time, the launch version is designed to refuse more advanced offensive cybersecurity requests, while still supporting defensive tasks such as secure code review and patching.

OpenAI is also trying to connect the product launch to its alignment story. The company says Astra is its most aligned model and describes new evaluations for whether a model stays within the authorized scope of a task. It says Astra showed better behavior than GPT-5.6 Sol in one test inspired by a prior security incident. The practical question will be how those safeguards perform at scale as more users delegate open-ended work to the model. Astra's debut signals that the next phase of frontier AI competition is moving toward agents that can operate tools, make judgments and complete complex tasks across real software environments.