One untuned prompt got GPT-5.6-Sol to match the best known OR algorithms.

The study tests inventory control, queueing network control and assortment optimization at two levels: solve one given instance, and return an algorithm mapping instance parameters to solutions before seeing any evaluation instance. With one untuned prompt and a Python sandbox on a fixed compute budget, gpt-5.6-sol matches or beats the best existing method on almost all evaluated instances, including the harder level-2 setting. Performance improves sharply across models released under eight months apart. The authors' conclusion is the useful one: treat a frontier LLM as a serious empirical baseline for algorithm design in well-specified OR problems, not as a novelty comparison. (arXiv)