Skip to main content

openai/gpt-realtime-2.1-mini/realtime

OpenAI’s GPT-Realtime-2.1 Mini is a distilled reasoning model for live voice agents: it listens and replies in streaming speech, with adjustable reasoning effort, function calling and a 128K-token window.

Text · Input $0.12Text · Output $0.48Audio · Input $2Audio · Output $4Image · Input $0.16per 1 million tokens80% below the publisher's list priceText · Cached input $0.012Audio · Cached input $0.06Image · Cached input $0.016OpenAI official priceText · Input $0.60Text · Output $2.40Audio · Input $10Audio · Output $20Image · Input $0.80

Workbench

Send a request

Quickstart

Examples use the selected endpoint’s public parameter definitions. Replace the API key and example media, then send the request from your server.

SDKs & integrations
POSThttps://api.spicyapi.ai/api/v1/jobs/createTask

API authentication

Keep the API key in the server environment variable SPICY_API_KEY. Use Bearer authentication and Content-Type: application/json for JSON requests.

Authorization: Bearer $SPICY_API_KEY
# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
  -H "Authorization: Bearer $SPICY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
  -d '{
  "model": "openai/gpt-realtime-2.1-mini/realtime",
  "input": {
    "prompt": "Calm ambient pads with a slow rising motif",
    "audio_input_format": "pcm",
    "audio_noise_reduction": "near_field"
  }
}'

# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
  -H "Authorization: Bearer $SPICY_API_KEY"

Generated from this model’s input schema — the same one the API validates every request against.

Input parameters

Place these fields inside the request body’s input object. Names and types match the API. Do not assume a default when none is specified.

realtime_sessionobjectOptional

Configure Mini’s advanced native function tools or turn detection here. Regular form controls must agree with any overlapping nested setting.

Default
Not specified
audio_noise_reductionstringOptional

Select Mini’s microphone denoising profile: near_field, far_field or none.

Default
"near_field"
Allowed values
"near_field""far_field""none"
audio_input_formatstringOptional

Mini expects mono 16-bit PCM at 24 kHz, or matching 8 kHz G.711 samples; encoded files are not accepted as raw chunks.

Default
"pcm"
Allowed values
"pcm""g711_ulaw""g711_alaw"
max_context_tokensintegerOptional

Set Mini’s retained history budget; history, a 4,096-token instruction allowance and this turn’s output must fit its 128,000-token window.

Default
4096
Constraints
minimum:512
maximum:123888
max_tokensintegerOptional

This Mini limit includes reasoning and audible output; an overly small value can stop the answer before a usable recording exists.

Default
512
Constraints
minimum:16
maximum:32000
speednumberOptional

Mini supports a speech-speed multiplier between 0.25 and 1.5; use a brief listening test before selecting a faster pace.

Default
1
Constraints
minimum:0.25
maximum:1.5
output_formatstringOptional

Mini streams PCM at 24 kHz or G.711 µ-law/A-law at 8 kHz; the task result wraps the selected encoding in WAV.

Default
Not specified
Allowed values
"pcm""g711_ulaw""g711_alaw"
reasoning_effortstringOptional

For Mini, start with minimal or low and raise the effort only when the spoken task needs additional reasoning.

Default
Not specified
Allowed values
"minimal""low""medium""high""xhigh"
voice_idstringOptional

Pick Mini’s speaking voice at the start of a conversation and keep it stable after audio generation begins.

Default
Not specified
Allowed values
"alloy""ash""ballad""coral""echo""sage""shimmer""verse""marin""cedar"
promptstringOptional

Use this text to start one GPT-Realtime-2.1 Mini response in the task interface; send microphone samples on a live WebSocket connection.

Default
Not specified
Constraints
minLength:1
maxLength:4096

Track the task

HTTP 202 means the task was accepted, not completed. Save the returned taskId and query that task. Do not create another task while waiting for its result.

Accepted response
HTTP/1.1 202 Accepted
Location: /api/v1/jobs/recordInfo?taskId=job_01k3m8x9q2z4v7n5p6r8s0t1w3

{
  "code": 200,
  "msg": "success",
  "request_id": "req_example_create",
  "data": {
    "taskId": "job_01k3m8x9q2z4v7n5p6r8s0t1w3",
    "state": "queued",
    "estimatedCost": "0.750000"
  }
}

The accepted response includes the estimate so your app can show the same amount in task history and billing.

Read the job
curl --get "https://api.spicyapi.ai/api/v1/jobs/recordInfo" \
  --data-urlencode "taskId=$TASK_ID" \
  -H "Authorization: Bearer $SPICY_API_KEY"

Start polling at about 2 seconds and gradually increase the interval. Stop at a terminal state. After a connection failure, query the original taskId. Retry creation only with the original idempotency key and body.

Or let the callback find you
import { verifyWebhook } from "@spicyapi/sdk";

// express.raw keeps the bytes we signed; express.json never matches.
app.post("/hooks/spicy", express.raw({ type: "application/json" }), (req, res) => {
  const { payload } = verifyWebhook({
    rawBody: req.body,
    timestamp: req.header("X-Webhook-Timestamp"),
    signature: req.header("X-Webhook-Signature"),
    payloadVersion: Number(req.header("X-Webhook-Payload-Version")),
    secret: process.env.SPICY_WEBHOOK_SECRET,
  });
  res.status(200).end();
  void handleFinishedTask(payload);
});

Pass callBackUrl when you create the task and the finished record is POSTed to you — no polling loop at all. Every delivery carries X-Webhook-Timestamp, X-Webhook-Signature and X-Webhook-Payload-Version. The signature is a base64 HMAC-SHA256 over taskId.timestamp.sha256(body), so it must be checked against the raw bytes: parsing the JSON and re-serialising it changes them and the check will fail. Failed deliveries are retried, so the same taskId can arrive twice — make the handler idempotent. Answer 2xx first and do the work afterwards; a slow handler is treated as a failed delivery.

Read the result

Consume output only when state is succeeded. On failure, use errorCode, errorMessage and request_id to investigate. Example costs and states do not represent an actual bill.

Completed response example
{
  "code": 200,
  "msg": "success",
  "request_id": "req_example_record",
  "data": {
    "taskId": "job_01k3m8x9q2z4v7n5p6r8s0t1w3",
    "model": "openai/gpt-realtime-2.1-mini/realtime",
    "state": "succeeded",
    "output": {
      "assets": [
        {
          "key": "tasks/2026/08/28/job_01k3m8x9q2z4v7n5p6r8s0t1w3/3f7a1c9e5b2d48f0a6c1e3d5b7f9a1c3.mp3",
          "mime": "audio/mpeg",
          "url": "https://example.r2.cloudflarestorage.com/results/job_01k3m8x9q2z4v7n5p6r8s0t1w3/result.mp3?X-Amz-Signature=SIGNATURE_FROM_RESPONSE",
          "expiresAt": "2026-08-28T12:20:08.000Z"
        }
      ]
    },
    "cost": "0.750000",
    "settled": true,
    "createdAt": "2026-08-28T12:00:00Z",
    "completedAt": "2026-08-28T12:00:30Z"
  }
}

For each available asset, open or download assets[].url directly without an API Authorization header. The signed URL expires at expiresAt; poll the task again to refresh it while the asset remains within its retention period. A pending or unavailable asset has no URL. Treat signed URLs as private.

SDKs & integrations

Use SpicyAPI in your app, terminal, AI assistant or visual workflow. Choose a setup guide below.

Use it in ComfyUI

Install the SpicyAPI nodes from ComfyUI Manager or with the command below, then search for GPT-Realtime-2.1 Mini · Realtime conversation on the canvas. The node has the same options as this page and shows its price before it runs.

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi