Claude's coding efficiency is largely due to their control plane governing the actual execution of the model - rules, evaluators and basically looping over results. Not related to the model training or fine tuning.
The models were already much better before Claude Code came out, let's not rewrite history here.