The model's native safety alignment (training-embedded weights) does not consume the context window, whereas system prompt injection at inference time actually consumes token quota and may conflict with user system prompts.