OpenAI Models Dominate Structured Code Edit Benchmark

I saw the same thing with Mailogy.

I only tested variants of gpt-3.5 and -4 but got ~50% invalid syntax errors with 3.5, and virtually none with 4.