$0.012 Either Way: Streaming Changes the Habit

01 / The $0.012 receipt has no streaming surcharge

Streaming changes when you see an answer, not the arithmetic used to price its tokens. If a streamed request and a non-streamed request have identical billable usage at identical rates, they cost the same. Receiving an answer in small pieces does not, by itself, make those pieces cheaper.

Here is a worked receipt using illustrative rates, not a CheapRelay price quote. Assume input costs $1 per million tokens and output costs $8 per million tokens. There are no cache reads, cache writes or other billable token categories in this example.

• ■ Input: 4,000 tokens × $1 / 1,000,000 = $0.004.
• ■ Output: 1,000 tokens × $8 / 1,000,000 = $0.008.
• ■ Total without streaming: $0.012.
• ■ Total with streaming and identical usage: $0.012.
• ■ Saving from changing only the delivery mode: $0.

For 10,000 requests with that exact usage, either delivery mode totals $120. That is arithmetic, not a workload forecast. Actual rates and token counts depend on the model, request and applicable billing categories.

02 / Stopping at 400 output tokens changes the receipt

The useful difference appears when a person acts on the partial answer. Suppose you can already tell that the response is heading in the wrong direction. Instead of waiting for all 1,000 output tokens, you press Stop.

If cancellation succeeds and billable output ends at 400 tokens, the revised receipt is:

• ■ Input: the same 4,000 tokens = $0.004.
• ■ Output: 400 tokens × $8 / 1,000,000 = $0.0032.
• ■ New total: $0.0072.
• ■ Difference: $0.0048 less, or 40% of the original total.

That saving comes from buying fewer output tokens, not from streaming. The input charge remains. Do not equate 400 tokens displayed in your interface with 400 tokens billed: generation can run ahead of display, and cancellation may not take effect immediately. Check the final usage record rather than estimating from the visible text.

03 / A second click can turn $0.012 into $0.024

Streaming can also change retry behaviour. A blank interface can make someone suspect that a request failed. Visible progress gives them evidence that work is happening, potentially reducing unnecessary duplicate submissions. That is a design opportunity, not a guaranteed reduction.

Under our original assumptions, two fully completed requests cost $0.024 rather than $0.012. Starting a replacement does not erase the first request's bill. If both continue generating, both can produce billable usage.

Build the interface accordingly: disable accidental double submission, show a clear running state, and connect Stop to actual upstream cancellation where supported. Merely hiding the response or stopping its animation is not the same as stopping generation.

04 / Longer conversations can consume the apparent saving

Streaming can make an assistant feel easier to use. That may encourage extra follow-up questions, longer sessions or more exploratory requests. None of those habits is necessarily bad, but they belong in the budget.

At the same illustrative input rate, sending another 4,000 input tokens adds $0.004 before any new output. If your application resends conversation history, later requests can carry more input than earlier ones. A cheaper individual response does not guarantee a cheaper session.

05 / Keep a usage receipt, not a streaming discount column

CheapRelay is pay-as-you-go, with no monthly fee or subscription. Billing is per real token, including cache reads and writes. For an actual estimate, use the applicable model rates and separate each billed category; do not apply the simple uncached receipt above to cached traffic unchanged.

Log request IDs, input usage, output usage, cache usage, cancellations and retries. Compare cost per completed task, not just cost per response. Sign up for CheapRelay when you are ready to measure those receipts against your own workload.

The practical rule is blunt: use streaming for visibility and control. Budget for the tokens actually billed. Any saving must come from changed usage, not from the appearance of moving text.