I'm using the Vercel AI SDK (ai package) to stream responses from a model, and I want to include extra provider metadata — specifically the cost — in the final finish metadata sent to the client.

Here’s the simplified version of my code:

1
const result = streamText({
2
    prompt,
3
    model: gateway(ai_model)}
4
);
5

6
const providerMetadataPromise = result.providerMetadata;
7

8
return result.toUIMessageStreamResponse({
9
  originalMessages: messages,
10
  sendReasoning: true,
11
  messageMetadata: ({ part }): Record<string, string> | undefined => {
12
    if (part.type === 'start') {
13
      return { model: ai_model };
14
    }
15

16
    if (part.type === 'finish') {
17
      let answerCost: string;
18
      providerMetadataPromise.then(data => {
19
        answerCost = (data!.gateway as any).cost;
20
      });
21
      return { model: ai_model, cost: answerCost! };
22
    }
23
  },
24
});
25

• If I use await result.providerMetadata before streaming starts, I get the cost correctly — but it blocks the start of the stream.

• If I remove the await and use .then(), streaming starts instantly — but the cost in the finish metadata is always undefined on the client.

• I also tried making messageMetadata async, but it seems the SDK doesn’t await it, so the cost never arrives.

I want:

• The streaming to start immediately (no delay).

• The final finish metadata to contain both { model, cost } values.

• The client to actually receive the cost along with the last message.

Is there a correct, non-blocking way to include async provider metadata (like cost) in the finish metadata for a streaming response in the Vercel AI SDK?