OpenAI launches GPT-6 Astra as flagship AGI-era model

GPT-6 Astra is the clearest signal yet that OpenAI is willing to say the quiet part out loud: it believes the industry has effectively crossed into an AGI phase. The launch is not just a bigger model, it is a full agentic stack designed to operate computers, browsers and specialized tools with far less human scaffolding. Benchmarks show meaningful jumps on long-horizon tasks, scientific workflows and cybersecurity, and Astra is explicitly positioned as a system that can discover zero-days and reverse engineer binaries. For the first time, OpenAI is labeling a public model “Critical” in cyber capability and shipping it with production misalignment monitoring.

Strategically, Astra pulls OpenAI back to the front of a race that has tightened as Anthropic, Google and Meta all shipped frontier-level systems. By focusing on agents and computer use, OpenAI is trying to own the layer where models turn into real work, not just chat. The flip side is that Astra appears harder to monitor: even OpenAI admits its written reasoning is less legible than GPT‑5.6 Sol’s under adversarial tests. That sets up a series of hard choices for labs and regulators about how much opacity they are willing to tolerate in exchange for capability. In practice, Astra will immediately reset expectations around what enterprise-grade AI agents should be able to do and will likely pull competitors into more aggressive agentic deployments.