How to access token usage metadata while using streamed response? [OpenAI, Anthropic]

[Sarkhad]

First of all, my goal is to implement chatbot with billing based on the token usage as the original API does.

It's known that OpenAI's and Anthropic's models return input-n-output tokens usage at the end of the streamed response.
Instead of implementing own low-level communication with LLM models, I'm aimed to integrate through Spring AI's abstraction.
Unfortunately, I've unsuccessfully tried high and low-level approaches that the framework provides to achieve the usage metadata. So, my question is how to get this metadata properly?

Here's some snippets if you wish:
` var messages = request.getMessages().stream()
.filter(m -> !m.role().equals("system"))
.map(m -> new AnthropicApi.RequestMessage(
List.of(new AnthropicApi.MediaContent((String) m.content())),
AnthropicApi.Role.valueOf(m.role().toUpperCase()))).toList();

    // var chatCompletionMessage = new OpenAiApi.ChatCompletionMessage("Hello world", OpenAiApi.ChatCompletionMessage.Role.USER);

    // Streaming request
    Flux<AnthropicApi.StreamResponse> streamResponse = anthropicApi.chatCompletionStream(
            new AnthropicApi.ChatCompletionRequest(
                    AnthropicApi.ChatModel.CLAUDE_3_OPUS.getValue(),
                    messages,
                    null,
                    request.getMaxTokens(),
                    0.8f,
                    true
            )
    );`

`var messages = request.getMessages().stream().map(m -> new OpenAiApi.ChatCompletionMessage(m.content(), OpenAiApi.ChatCompletionMessage.Role.valueOf(m.role().toUpperCase()))).toList();

    // var chatCompletionMessage = new OpenAiApi.ChatCompletionMessage("Hello world", OpenAiApi.ChatCompletionMessage.Role.USER);

    // Streaming request
    Flux<ChatCompletionChunk> streamResponse = openAiApi.chatCompletionStream(
            new OpenAiApi.ChatCompletionRequest(messages, request.getModel(), (float) request.getTemperature(), true));
    //Prompt prompt = new Prompt(new UserMessage(message));
    //return chatModel.stream(prompt);

    return streamResponse;`

` var openAiChatOptions = OpenAiChatOptions.builder()
.withModel(request.getModel())
.withTemperature((float) request.getTemperature())
.withMaxTokens(request.getMaxTokens())
.build();
var chatModel = new OpenAiChatModel(openAiApi, openAiChatOptions);

    var messages = request.getMessages().stream().map(m -> {
        switch (MessageType.fromValue(m.role())) {
            case USER -> {
                return (Message) new UserMessage((String) m.content());
            }
            case ASSISTANT -> {
                return (Message) new AssistantMessage((String) m.content());
            }
            case SYSTEM -> {
                return (Message) new SystemMessage((String) m.content());
            }
            case FUNCTION -> {
                return (Message) new FunctionMessage((String) m.content());
            }
            default -> throw new IllegalStateException("Unexpected value: " + MessageType.fromValue(m.role()));
        }
        //new Message(m.content(), OpenAiApi.ChatCompletionMessage.Role.valueOf(m.role().toUpperCase()))
    }).toList();


    Flux<ChatResponse> response = chatModel.stream(
            new Prompt(messages));
    return response; //.map(ChatResponse::getResult);`