Streaming and batch cost the same per token — the real difference is in latency, per-request overhead, and retry behaviour. Here's the real math, a worked example, and a break-even formula to decide for your own workload.