Per-token rates are usually the same
Most major APIs bill input and output tokens the same whether you stream tokens or wait for a full JSON blob. If your comparison only multiplies tokens × price, streaming and non-streaming look identical.
Early cancel is the real API lever
In chat UIs, users often stop generation mid-stream. Streaming lets you halt decode and avoid paying for the remaining output. Non-streaming has already paid for the full completion by the time the user sees it.
TTFT is a product cost, not a line item
Streaming improves time-to-first-token even when the API invoice matches. If you assign a wait cost (SLA, support load, or conversion), that UX gap shows up as effective savings—pair this with the latency planner for model-to-model tradeoffs.
FAQ
Is batch API streaming?
No. Batch is an offline, delayed completion path with discounted rates. Streaming is about delivery of a realtime response.