OpenAI launches GPT-6 Astra, its most advanced model yet

GPT-6 Astra arrives as the successor to GPT-5.6 Sol, and OpenAI positions it as a major step forward by pairing raw intelligence with improved judgment. One of the clearest signs is its performance on complex benchmarks: a 99.9% score on ARC-AGI-3 (abstract reasoning) and a 95.9% score on BenchCAD (converting 2D views into 3D CAD code) illustrate the breadth of the gains versus its predecessor. The model's ability to handle longer, more complex tasks is also highlighted by its performance on the Agents' Last Exam and OSWorld 2.0, where it uses roughly 47% less time per task than GPT-5.6 Sol.

Much of the surrounding context focuses on responsibility. OpenAI's claim that Astra is its most aligned model yet is supported by internal tests showing it is less likely to misrepresent its capabilities, follow instructions beyond its scope, or attempt to bypass system safeguards. This effort addresses a key challenge: as AI agents gain autonomy in software and computer-use tasks, ensuring they stay within authorized boundaries becomes as critical as their raw power.

The strategic stakes are significant. While Astra's ability to identify and develop zero-day exploits is framed as a defensive tool for finding and patching weaknesses, it also raises the risk of misuse. OpenAI is balancing this by refusing some advanced cyber tasks at launch while planning to expand access through OpenAI Daybreak. The central question for the coming weeks and months is likely whether its alignment and monitoring systems hold up in real-world deployments, especially as safeguards are selectively relaxed for defensive work.