Codex Outperforms DRL in Fluid Control With Handwritten Code, $14 Total Cost

According to Beating, OpenAI post-training researcher Paul Garnier demonstrated that Codex 5.5 can generate interpretable control code outperforming deep reinforcement learning baselines in fluid mechanics applications. Rather than training neural networks, Garnier used the model to iteratively refine Python scripts by analyzing physics simulations, achieving superior performance in over half of tested scenarios.

The AI-generated control rules were physically interpretable, such as "delay jet injection when local curvature exceeds threshold." Unlike neural network black boxes, the code-based approach proved robust under distribution shifts; when test duration was extended four-fold, traditional DRL models collapsed while the physics-informed code remained stable. Implementing the full control strategy consumed 21.25 million tokens, totaling under $14.