New study reveals Python cheapest for AI coding agents, Rust most expensive

A new arXiv paper (2607.22807) from Wu, Anderson, and Guha investigates how coding agents consume tokens differently across programming languages. Testing five recent models on problems in Python, Java, Rust, and OCaml, they controlled for difficulty and found stark, consistent variation in token consumption. Rust, for instance, demanded far more tokens than Python. The cost difference is large enough that language choice alone can dramatically affect the budget for AI-assisted development.

Digging into agent trajectories, the researchers uncovered inefficient behaviors: agents repeatedly produce noncompiling solutions in unfamiliar languages and revise solutions that already pass. They also plan in code comments, distrust provided tests in favor of self-invented inputs, and sidestep unfamiliar target languages by prototyping in Python. These results suggest that token efficiency by language should be a standard metric when benchmarking and developing multilingual agents, and serve as a guide for 'tokenmaxxers' seeking the most expensive language to avoid.