OpenAI Launches 'Computer-Using' GPT-6 Astra, Declares Onset of AGI Era

OpenAI has officially launched GPT-6 Astra, an agentic AI system designed to directly control digital interfaces and complete complex, multi-step tasks across professional, scientific, and software engineering domains.

The launch reportedly marks a fundamental transition from standard conversational chatbots to autonomous execution engines. Featuring a 1.05-million-token context window and up to 128,000 output tokens, Astra operates computer software just like a human interacting with forms, managing calendars, updating CRM databases, running tests, and generating complex documentation without direct user intervention.

Understanding AGI 

AGI stands for Artificial General Intelligence. Unlike narrow AI which excels only at specific, pre-defined tasks like translation or image recognition AGI refers to autonomous systems that possess human-level intellectual capabilities across a broad spectrum of cognitive, creative, and problem-solving tasks.

Technical Milestones & Performance Benchmarks

Astra set a series of baseline performance records across automated reasoning, mathematics, and agentic desktop execution:

• Reasoning & Math: Scored 99.9% on the ARC-AGI reasoning test and 98% on FrontierMath Tier 4.
• Computer Navigation: Achieved 72.6% on the OSWorld 2.0 desktop task benchmark, outperforming GPT-5.6 Sol (65.7%).
• Speed Efficiency: Reduced task completion time down to 40 minutes from Sol’s 75-minute average, a 47% drop in execution duration.
• Cybersecurity: Achieved a perfect 100% score on the ExploitBench security assessment. Early enterprise testing by legal-tech firm Legora demonstrated a 40% performance boost in financial statement processing.

Declaring the AGI Era

OpenAI President Greg Brockman framed the rollout as a qualitative turning point in tech history, declaring, "I think it’s not unreasonable to feel that we are now in the AGI era." He noted that the term is no longer tethered to contractual limits, asserting that AI has crossed the boundary from assistant to autonomous partner.

Safety Alignment & Critical Cybersecurity Risks

Because Astra can discover vulnerabilities and write functional code exploits, OpenAI  reportedly assigned it a "Critical" cybersecurity threat rating. However, safety evaluations showed improved boundary adherence: while un-safeguarded tests on GPT-5.6 Sol resulted in authorized target breaches 48% of the time, Astra maintained strict operational boundaries across every test case.cessor model GPT-5.6 Sol breached authorized task targets 48% of the time, whereas Astra maintained set boundaries in every test case.