Is there any technical analysis of why contex grows slower in Pi compared to codex and CC?

different system prompts, codex will have system prompt telling the model to gather a lot more context before starting work