OpenAI Launches GPT-6 Astra: Frontier Model for Autonomous Computer Use and Enterprise Engineering

On September 11, 2026, OpenAI officially launched GPT-6 Astra, its most capable and aligned frontier foundation model to date. Now accessible across ChatGPT Work, Codex, and the OpenAI API, GPT-6 Astra establishes new state-of-the-art benchmarks in autonomous computer use, terminal-based software engineering, scientific research, and enterprise data reasoning, allowing organizations to execute complex multi-step workflows across standard desktop applications without custom API integrations.

A New Paradigm in Autonomous Computer Use

Unlike predecessors that required pre-structured API scaffolding, GPT-6 Astra interacts directly with operating system graphical user interfaces (GUIs), terminal consoles, and web browsers. In benchmark evaluations from the Financial Modeling World Cup, Astra completed complex multi-tab Excel financial modeling challenges four times faster than human champions. The model operates native applications—such as Figma, Power BI, and Oracle Analytics—by perceiving UI pixels, understanding visual intent, and issuing native mouse and keyboard actions.

Benchmark Breakthroughs & Efficiency Metrics

• Terminal-Bench 4.0: Scores a record 57.9% accuracy on complex terminal engineering tasks (surpassing GPT-5.6 Sol at 37.3% and Claude Fable 5.1 at 55.8%).
• DeepSWE v1.1 Software Engineering: Sets a world-record 74% solve rate on real-world multi-file pull requests while catching 20% more bugs.
• Cost Efficiency: Completes complex agentic tasks with fewer tokens and retries, delivering a 9% to 63% lower cost per completed task compared to competing frontier models.
• API Pricing: Set at $10.00 per million input tokens and $50.00 per million output tokens with Zero Data Retention support for enterprise endpoints.

Frontier Alignment & Cybersecurity Governance

Astra is the first model to reach the 'Critical' cybersecurity capability threshold under OpenAI's internal Preparedness Framework. To ensure enterprise security, Astra incorporates fine-grained admin controls, human-in-the-loop authorization gates for consequential actions, and produces unintended tool outputs 89% less often than GPT-5.6 Sol.

Model Profile Sheet