Study new fields

Handles difficult long-running tasks more effectively by investigating problems, using tools, recovering from mistakes, and verifying results in realistic environments. Extends beyond software engineering into data science, finance, legal work, and other computer-based knowledge tasks with a broader training mix. Improves agent-style computer work through training on large volumes of codebase interactions and developer-agent behavior with software tools. Gains stronger robustness on hard problems through reinforcement learning environments designed specifically to remain challenging even for frontier systems. Offers a separate fast variant, indicating a higher-speed option alongside the main weight class for lower-latency workflows.