Most of the performance on coding comes from RL, not distillation.

Distillation helps with world knowledge and things like that.

They're not distilled. Stop spreading anthropics misuse of the term.

They do use it for synthetic data/judging though, so yes, hard to outperform.

Not that they need to. If they can basically match it for a fifth of the price.

Is that true? If the distillation is not lossy and the model runs much faster due to less resource consumption, then it may outperform.