# Responses and usage

Source: https://docs.getworg.com/responses

A response has the OpenAI shape. usage says what the request used, and that is what Worg bills.

## Response

Response:

```json
{
  "id": "chatcmpl-req_tVJMBaXIcyDHuiQT2Ljh8WAG",
  "object": "chat.completion",
  "created": 1793729613,
  "model": "deepseek-v4.1-flash:cached",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "Halo!" },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 9,
    "completion_tokens": 4,
    "total_tokens": 13,
    "prompt_tokens_details": { "cached_tokens": 0 }
  }
}
```

- `id` is `chatcmpl-` followed by the request ID, the same value as the `x-request-id` header. In a stream every chunk carries the same `id`.
- `model` is the lane that answered. When you sent a renamed ID, it is the current ID. Through a combo, it is the lane that served the request.
- `message` has `content`, `reasoning_content` when the model reasoned (see [Reasoning](https://docs.getworg.com/reasoning)), and `tool_calls` when it called a function.
- `logprobs` is `null` unless you asked for it on a lane that returns it (see [lane differences](https://docs.getworg.com/models#differences)).

## Finish reasons

- `stop`: the model finished, or reached one of your `stop` strings.
- `length`: the output reached `max_tokens`. The text is cut off, so JSON output may not parse.
- `tool_calls`: the model wants your code to run the functions in `tool_calls`.

## Token usage

- `prompt_tokens` counts all input: messages, tool definitions, and images.
- `prompt_tokens_details.cached_tokens` is the part of the input served from the model's cache. It is billed at the cached rate. Cache hits are not guaranteed; they come when a request repeats the start of a recent one.
- `completion_tokens` counts all output, reasoning included. Reasoning tokens are not reported separately, because not every lane reports them.
- With reasoning on, `prompt_tokens` is a little higher for the same messages (16 tokens became 42 in one test), because the model's template adds instructions.

## How a request is billed

Worg multiplies each kind of token by the lane's price per 1M tokens and rounds the sum up to a whole micro-dollar (US$0.000001). New input is`prompt_tokens` minus `cached_tokens`. The price is the one in force when the request ran, and it is recorded with the request.

Example: a request on `deepseek-v4.1-flash:cached` with a long document in the system message, sent a second time a few seconds after the first. 5,774 of its 6,277 prompt tokens came from the cache.

| Part | Tokens | US$ per 1M | Micro-US$ |
| --- | --- | --- | --- |
| New input | 503 | 0.3 | 150.9 |
| Cached input | 5,774 | 0.006 | 34.644 |
| Output | 2 | 1.2 | 2.4 |
| Charged |  |  | 188 (US$0.000188) |

Prices read from the API when this page was built. Current prices: [Models and pricing](https://getworg.com/models/).

- Requests refused before a lane is called, and requests the model rejects, are not billed.
- If a lane sends no usage, Worg bills an estimate: the input at the bound used for the [cost hold](https://docs.getworg.com/rate-limits#hold), and one output token per 3 bytes of generated text.
- Every request appears in the console under Usage with its tokens and cost.

## Headers

- `x-request-id` on every chat completion response, errors included, except `401` and `413 body_too_large`. Send it when you contact support.
- `retry-after`, in seconds, on `429` responses. See [Retries](https://docs.getworg.com/errors#retries).

Next: [Streaming](https://docs.getworg.com/streaming) · [Rate limits](https://docs.getworg.com/rate-limits)