I spent 1 week and hundreds of dollars to run a real-world programming benchmark (real project back to back) through 30 LLMs to assert which is best and why.