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.