To be fair Codex was much better than GPT-3 on reasoning benchmarks like MMLU and the like. And people have kind of noticed the Code trained models to reason better. Don't know if a paper was published about that though.