My own speculation is that, in addition to even more scaling up, Anthropic has been able to better optimize the core LLM for its use inside a highly optimized harness, due to availability of all the data Anthropic has been able to collect in its surge in popularity in the last months.