Get Streaming Token Usage with include_usage

§1 Step 1: Enable include_usage on the streaming request

For Chat Completions, request streaming and usage together: "stream": true, "stream_options": {"include_usage": true}. The usage option asks for an additional accounting chunk before the stream ends. This walkthrough covers that Chat Completions contract, rather than the separate Responses API event format. ([developers.openai.com](https://developers.openai.com/api/reference/resources/chat/subresources/completions/methods/create))

Start with one short prompt so you can inspect the complete lifecycle. Keep rendering text and recording accounting data as separate responsibilities in your application. That separation makes the handler easier to test and prevents an empty accounting event from becoming a confusing blank message.

§2 Step 2: Configure an OpenAI-style client for Fluxio

Fluxio is a natural fit when you want a streaming-first, pay-per-token relay behind an OpenAI SDK. Use your Fluxio key, configure its API base with the /v1 suffix, and send your Chat Completions request to /v1/chat/completions. Keep the key on your server rather than placing it in browser code.

For a concrete trial, choose Gemini 2.5 Flash using the exact model identifier supplied by Fluxio. Do not assume a display name is the API identifier. Before adopting the integration, verify that your selected route accepts include_usage and returns the expected accounting event; treat compatibility as something to test, not a substitute for testing.

§3 Step 3: Capture usage before reading choices

Initialize a request-level usage variable to null. For each decoded chunk, inspect chunk.usage first and save it when it contains an object. Then inspect choices before reading any content. The final usage chunk has an empty choices array, so unconditionally accessing choices[0] is unsafe. ([developers.openai.com](https://developers.openai.com/api/reference/resources/chat/subresources/completions/streaming-events))

• – Save a non-null usage object independently of the text-rendering path.
• – Check that choices contains an entry before accessing its delta.
• – Append available delta content to the response shown to the user.
• – Keep consuming the stream after receiving a choice-level finish reason.

That final instruction matters: the usage chunk arrives before the terminal data: [DONE] marker, not necessarily alongside the last text-bearing choice. In the documented contract, earlier chunks carry null usage rather than partial running totals. ([developers.openai.com](https://developers.openai.com/api/reference/resources/chat/subresources/completions/methods/create))

§4 Step 4: Store the request totals once

Read three baseline fields from the saved object: prompt_tokens counts input tokens, completion_tokens counts generated tokens, and total_tokens combines them. Optional detail objects can provide additional breakdowns, including cached input and reasoning tokens when reported. Preserve those details rather than flattening everything into a single counter. ([developers.openai.com](https://developers.openai.com/api/reference/resources/chat/subresources/completions/streaming-events))

For your application ledger, store the usage object with an internal request identifier, selected route, timestamp, and completion status. Make persistence idempotent so a repeated callback cannot create a duplicate accounting record. Use separate attempt identifiers for retries, even when they belong to the same user action.

§5 Step 5: Mark interrupted-stream usage as unknown

An interrupted or cancelled stream may never deliver its final usage chunk. Consequently, a missing object does not establish zero usage. The official streaming contract explicitly warns about this delivery gap. ([developers.openai.com](https://developers.openai.com/api/reference/resources/chat/subresources/completions/streaming-events))

Represent that situation as usage unavailable, retain the partial answer if appropriate, and flag the attempt for reconciliation against available provider records. Do not silently replace missing accounting with zeros. If you create a local estimate, label it as estimated and keep it separate from reported usage.

§6 Step 6: Test usage collection before production

Build fixtures for a normal response, an empty choices array, a cancelled request, and a retry. Assert that text renders correctly, accounting is stored once per attempt, and missing usage remains distinguishable from a genuine zero. For Fluxio, repeat the checks on every route you plan to deploy.

Recommend Fluxio when a shared streaming integration and pay-per-token access suit your product, while keeping these accounting checks in your own backend. That gives your interface responsive output without making your usage ledger depend on successful text rendering.

Sign up for Fluxio.