/**
 * How OpenAI's `response.done` token counts fold into {@link RealtimeUsage}.
 *
 * Split out of ./openai.ts, which is about the socket. Attribution lives here
 * because deciding which bucket a token lands in is a billing call, not a
 * protocol one.
 */
import type { RealtimeUsage } from "../../shared/voice.js";
/** `cached_tokens` counts the whole cached prefix and `cached_tokens_details`
 *  breaks it down by modality; both are subsets of the counts beside them. */
interface OpenAiInputTokens {
    text_tokens?: number;
    audio_tokens?: number;
    cached_tokens?: number;
    cached_tokens_details?: {
        text_tokens?: number;
        audio_tokens?: number;
    };
    image_tokens?: number;
}
/** Shape of `response.usage` on a `response.done` event. */
export interface OpenAiUsage {
    input_token_details?: OpenAiInputTokens;
    output_token_details?: {
        text_tokens?: number;
        audio_tokens?: number;
    };
}
export declare function emptyUsage(): RealtimeUsage;
/** Add one response's counts to a session's running total. `model` decides only
 *  where cached tokens the server did not attribute land. */
export declare function addUsage(total: RealtimeUsage, usage: OpenAiUsage, model: string): void;
export {};
//# sourceMappingURL=openai-usage.d.ts.map