GPT-6 Astra Operational

GPT-6 Astra is OpenAI's flagship large language model, officially described as "the world's most intelligent model and the one most aligned with human intent." It goes beyond traditional chat, evolving into an AI agent that can independently operate software, browse the web, and complete complex tasks. It set records in tests including a 99.9% score on ARC-AGI-3 and a 100% score on ExploitBench, and can handle professional work in programming, scientific research, and cybersecurity.

Computer Use: Independently operates computer software, browses the web, fills out forms, updates CRMs, and organizes schedules.
Software Engineering: Handles complex programming and codebase comprehension, achieving SOTA on both Terminal-Bench and DeepSWE.
Scientific Research: Directly analyzes experimental data and generates charts, scoring 97.6% on FrontierMath Tier 4.
Cybersecurity: Discovers and exploits zero-day vulnerabilities, achieving 100% on ExploitBench.
3D and Hardware Design: Operates Blender, Unity, and KiCad to create 3D models, game environments, and PCB designs.
Professional Documents: Generates beautifully laid-out posters, presentations, and ready-to-use business documents.

Large-Scale Pretraining: Pretraining was conducted for the first time at the U.S. Stargate Texas facility using more than 100,000 GPUs, making it one of OpenAI's largest training projects to date.
Model Alignment and Safety: Combines years of reinforcement learning and alignment research, using previous models for supervision; in boundary-breaking task tests, the unauthorized action rate fell from 48% to 0%.
Long-Horizon Context Management: Codex saves notes across context windows and supports searching past messages and tool outputs, ensuring that long-running tasks retain critical information.
Tiered Safety Policies: The standard version refuses advanced cybersecurity requests, while vetted security teams can access more open capabilities through the Daybreak program, balancing performance and risk.

Get Access: Subscribe to ChatGPT Plus/Pro/Business/Enterprise or apply for API access to obtain permission to use GPT-6 Astra.
Switch Models: Select GPT-6 Astra as the current model in the ChatGPT client or API endpoint.
Enter Instructions: Use natural language to provide complex task instructions, supporting scenarios such as programming, design, scientific research, and office automation.
Enable Computer Use: Turn on the Computer Use feature and authorize Astra to operate software, browse the web, and call external tools autonomously.
Confirm Interactions: Provide real-time feedback based on execution progress, confirm key steps, or adjust requirements until the task is complete and the results are exported.

Transformative Agent Capabilities: Evolves from chat-based question answering into an AI agent that can independently complete complex tasks, directly operate software, browse the web, and handle professional workflows.
Top Benchmark Scores: ARC-AGI-3 reached 99.9%, ExploitBench reached 100%, and FrontierMath Tier 4 reached 97.6%, setting new records across multiple tests.
Software Engineering Leadership: Achieves SOTA on both Terminal-Bench and DeepSWE, requires fewer code-generation iterations, and retains long-horizon task context.
Scientific Research Capabilities: Can directly access professional software to analyze experimental data, participate in mathematical research, and advance work on the prime-gap problem.
Creative and Hardware Design: Supports Blender, Unity, and KiCad for 3D modeling, game environment construction, and PCB design.
Cybersecurity as a Double-Edged Sword: Can discover zero-day vulnerabilities, but the Daybreak tiered strategy strictly controls the scope of access.

Project website: https://openai.com/index/gpt-6-astra/

Software Engineering: Autonomously writes and debugs complex code and performs database migrations, achieving industry-leading results on both Terminal-Bench and DeepSWE.
Scientific Research: Directly accesses professional software to analyze gene sequencing data, generate experimental charts, and contribute to research on advanced mathematical problems.
Office Automation: Independently operates tools such as Excel and Power BI to complete form filling, CRM updates, schedule organization, and complex data analysis.
3D and Hardware Design: Creates 3D models and builds game environments in Blender and Unity, or uses KiCad to design PCB layouts.
Cybersecurity: Proactively scans for and discovers zero-day vulnerabilities, helping enterprises fix system defects while using tiered policies to prevent abuse.