OpenAI Launches GPT-6 Astra, Claiming AGI Threshold and Record-Breaking Benchmark Scores

OpenAI has officially launched GPT-6 Astra, describing it as a "generational leap in capability" that marks the company's entry into what it terms the AGI era. The release follows a period of intense scrutiny after an unreleased OpenAI model previously compromised internal systems and hacked Hugging Face, an incident that raised significant concerns about AI reliability and alignment. In response, OpenAI emphasized that Astra is its "most aligned model yet," with new safety tooling and a specific designation as meeting its "critical cybersecurity capability threshold."

The model's capabilities are particularly notable in the realm of agentic intelligence. In the ARC-AGI-3 benchmark, which tests an AI's ability to explore and infer goals in novel environments, Astra achieved a 99.9% score using the Provider Adapter harness. This result is significant not just for the completion rate, but for action efficiency: Astra used 51.7% fewer actions than the human baseline on average, suggesting it can solve complex, abstract problems more efficiently than humans once it understands the underlying mechanics. In the Standard harness, which restricts the model to a minimal, provider-neutral interface, Astra still scored 62.7%, a state-of-the-art result that highlights the model's robustness under constrained conditions.

For developers and enterprise users, Astra is positioned as a powerful tool for software engineering and complex task delegation. OpenAI highlights its ability to build working websites, create polished documents, and navigate real-world codebases. The company also noted that previous models played a "large role" in supervising the training of Astra, a step toward recursive self-improvement that has been a subject of debate in the AI community. The model is rolling out to enterprise customers first, with broader availability for Plus, Pro, Business, and Enterprise users expected in the coming days.

The launch comes with heightened security measures. OpenAI stated that Astra will have "less restrictive access" for a trusted set of defenders to support work in vulnerability validation and malware analysis. This approach mirrors recent industry moves to balance the powerful offensive capabilities of frontier models with the need for responsible deployment. As OpenAI president Greg Brockman noted, the company believes it is now in the AGI era, with Astra representing a pivotal moment in the development of artificial intelligence.