Why should we assume that LLMs would be stagnant at current levels when, so far, they haven't stopped improving? I remember when simply using Copilot to just auto-complete 1 line of code was a groundbreaking unimaginable advancement. That was only a few years ago. New and improved models are being released nearly every week. The open-source models are nearly as advanced as the closed source models (if you have the hardware to run them at full capacity).
Later, you'll realize that there are no new senior coders to burn out, because you got rid of the junior coders, but we're not there yet.