Introduction
GPT-Realtime-2.1 Mini is a distilled reasoning voice model for applications where a spoken exchange should remain brief and responsive. It listens to audio, retains conversation context and generates speech directly. Mini can also receive text or images within the native conversation protocol. The offered connection uses server WebSocket transport, with a separate single-response task for trying a short written request.
Key features & highlights
Mini’s positioning favors frequent, lower-cost voice interactions rather than long spoken monologues. Reasoning effort remains adjustable, so a simple confirmation can use a small budget while a turn that requires checking several conditions can receive more processing. Test the actual interaction rhythm with the chosen instructions; a higher effort setting can increase the wait before useful speech.
GPT-Realtime-2.1 Mini includes function calling, preset voices and streaming text or audio responses. Its model window is 128,000 tokens, and output can reach 32,000 tokens within that total. A connection reserves a caller-selected history allowance and a per-response output limit, allowing an application to keep short sessions economical without treating every open connection as a request for more speech.
Request parameters
| Parameter | Required | Description |
|---|---|---|
prompt | No | Use this text to start one GPT-Realtime-2.1 Mini response in the task interface; send microphone samples on a live WebSocket connection. at least 1 characters · at most 4096 characters |
voice_id | No | Pick Mini’s speaking voice at the start of a conversation and keep it stable after audio generation begins. |
reasoning_effort | No | For Mini, start with minimal or low and raise the effort only when the spoken task needs additional reasoning. |
output_format | No | 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. |
speed | No | Mini supports a speech-speed multiplier between 0.25 and 1.5; use a brief listening test before selecting a faster pace. |
max_tokens | No | This Mini limit includes reasoning and audible output; an overly small value can stop the answer before a usable recording exists. |
max_context_tokens | No | 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. |
audio_input_format | No | Mini expects mono 16-bit PCM at 24 kHz, or matching 8 kHz G.711 samples; encoded files are not accepted as raw chunks. |
audio_noise_reduction | No | Select Mini’s microphone denoising profile: |
realtime_session | No | Configure Mini’s advanced native function tools or turn detection here. Regular form controls must agree with any overlapping nested setting. |
Pricing
Text · Input $0.12 / Text · Output $0.48 / Audio · Input $2 / Audio · Output $4 / Image · Input $0.16 per 1 million tokens
Use cases
For a voice-enabled checklist using GPT-Realtime-2.1 Mini, let the listener answer one item at a time. Store the confirmed result in the application before the model asks the next question, so an interruption does not silently advance the list.
For a compact spoken help widget, provide a small set of product facts and functions. Have Mini answer the immediate question, then ask whether more detail is useful instead of reciting an entire help article.
Prompt guide & pro tips
Specify a short response style for GPT-Realtime-2.1 Mini, including a preferred language and a limit such as one question per turn. Name the function that retrieves a fact and require its result before the assistant reports completion. Avoid embedding sample conversations that accidentally instruct the model to narrate internal steps; review spoken output as well as any accompanying transcript.
Begin with minimal or low reasoning and an output limit that can complete the expected answer. For microphone input, choose the exact sample encoding and test voice detection with pauses typical of your audience. The application must stop or flush its audio player when the user interrupts. Mini’s standalone response task returns a WAV file; live PCM or G.711 chunks need client-side playback. Independent speech recognition uses a transcription connection, so a voice response does not silently add a second transcription bill.
API quick start
# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${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"
See the API reference for authentication, the complete schema, polling, and error handling.

















