Online RL is reinforcement learning where the policy generates new data during training and is updated on that fresh data — higher compute cost but the policy improves on its own distribution. This approach ensures the model learns from its current behavior, leading to more effective exploration and potentially better final performance compared to offline methods.