Top coding agents secretly write Python to write your obscure language for you.

A June 9 paper finds frontier agents like Claude Opus 4.6 and GPT-5.4 tackle esoteric or unfamiliar languages not by coding in them directly but by writing Python that generates the target-language code (arXiv 2606.10933). Forbidding this metaprogramming caused large performance drops in the strongest agents, while feeding Opus-derived helper code sharply improved weaker models and barely moved Haiku 4.5. The takeaway: strategy quality, not raw model size, increasingly differentiates agent capability on hard tasks. The good model isn't just smarter, it has better tactics. That's a buildable insight, you can hand those tactics to cheaper models.