GPT-Realtime-2.1
OpenAI's current speech-to-speech reasoning model for voice agents on the Realtime API
Context window
128K
Input / 1M tokens
$32.00
Output / 1M tokens
$64.00
Provider
OpenAI
Per 1M tokens: audio input $32.00, cached audio input $0.40, audio output $64.00; text input $4.00, cached $0.40, output $24.00; image input $5.00 ($0.50 cached). The listed $32/$64 is audio in and audio out. GPT-Realtime-2.1 Mini costs $10/$20 for audio and $0.60/$2.40 for text. Max output is 32K tokens and the knowledge cutoff is September 30, 2024. · Data verified 2026-11-11
GPT-Realtime-2.1 (July 6, 2026) is OpenAI's current model for the Realtime API: a speech-to-speech reasoning model with configurable reasoning effort, function calling and image input, aimed at production voice agents. It updates GPT-Realtime-2 with better recognition of letters and numbers, better handling of silence and background noise, and better interruption behavior. It has a 128K context and up to 32K output tokens. A cheaper GPT-Realtime-2.1 Mini launched alongside it, and OpenAI said it cut p95 latency by at least 25% across its Realtime voice models through improved caching. GPT-Live, OpenAI's full-duplex voice model, uses a separate Live endpoint.
How to access it
OpenAI API through the Realtime endpoint (v1/realtime) only; it is not available in Chat Completions, Responses or Batch. Model IDs gpt-realtime-2.1 and gpt-realtime-2.1-mini.
Strengths
- ✓Configurable reasoning effort for harder voice tasks
- ✓Function calling and image input inside a live voice session
- ✓Better alphanumeric recognition and noise handling than GPT-Realtime-2
- ✓A Mini version for lower cost and latency
Best for developers who...
When to choose it (and when not to)
Reach for GPT-Realtime-2.1 when...
- →You are building a voice agent that must call tools and follow instructions in real time
- →You already use GPT-Realtime-2 and want its direct update
- →You want to trade cost for quality with the Mini version
Look elsewhere if...
- ✕You only need transcripts (GPT-Transcribe and GPT-Live-Transcribe cost far less per minute)
- ✕You want simultaneous listening and speaking with background delegation (GPT-Live)
- ✕You need knowledge after September 2024 without tools
How to use it
- ›Raise reasoning effort only for hard turns; higher effort adds latency and output tokens
- ›Keep tool schemas short and specific so calls fire reliably during live audio
- ›Keep long system instructions fixed so they are billed at the cached rate
Quickstart
Pythonimport asyncio
from openai import AsyncOpenAI
async def main():
client = AsyncOpenAI()
async with client.realtime.connect(model="gpt-realtime-2.1") as connection:
await connection.session.update(session={"type": "realtime", "output_modalities": ["text"]})
await connection.conversation.item.create(
item={
"type": "message",
"role": "user",
"content": [{"type": "input_text", "text": "Say hello!"}],
}
)
await connection.response.create()
async for event in connection:
if event.type == "response.output_text.delta":
print(event.delta, flush=True, end="")
elif event.type == "response.done":
break
asyncio.run(main())Install with pip install openai. Text-only version of the Realtime example in the openai-python README; for audio, add audio output modalities and stream microphone input.
API model id: gpt-realtime-2.1
Compare GPT-Realtime-2.1
Compare GPT-Realtime-2.1 with any other model
Build a comparison →All model comparisons →