OpenAI launches GPT-6 Astra, its most advanced model

GPT-6 Astra is the culmination of OpenAI's work across pre-training, reinforcement learning, and alignment, bringing together advances in raw intelligence with a new focus on safely acting in the world. The model is positioned as a major step in computer use, handling tasks like filling out forms, updating CRM records, and running frontend QA checks, and its efficiency gains are notable: it achieves higher performance in about 47% less time per task than GPT-5.6 Sol on OSWorld 2.0. OpenAI also highlighted GPT-6 Astra's role in scientific discovery, sharing two new results on prime number gaps, including helping establish a stronger bound of 186.

The launch is paired with an unusual emphasis on safety and alignment. The model is described as OpenAI's most aligned, with improvements in respecting task boundaries and communicating transparently. However, OpenAI also acknowledges a significant increase in cyber capabilities, meeting the Critical threshold in cybersecurity, and the new model can identify zero-day exploits. The company is deploying misalignment monitoring in production and is being careful to make the deployment safe, with extra safety checks that can slow or pause work. Expert-led assessments found Astra could achieve arbitrary code execution in hardened browsers when run without safeguards.

The immediate stakes for OpenAI are whether it can successfully navigate the tension between delivering a more capable and efficient tool and managing the risks of its advanced cyber abilities. The version launching today refuses more advanced cybersecurity tasks, but OpenAI's plan to roll out less restrictive safeguards through OpenAI Daybreak presents a potential inflection point. The model's real-world impact will likely be shaped by how its alignment holds up in complex, delegated tasks and how its cyber capabilities are managed as access expands.