Yes, and the gap is larger than most expect. A 4K system prompt can add hundreds of ms to TTFT before the model says anything. Streaming hides decode latency but does nothing for prefill. Trimming the system prompt usually beats any decode-side optimization people try first.