OpenAI Launches Realtime 2.1 API with Configurable Voice Reasoning
OpenAI released gpt-realtime-2.1 and gpt-realtime-2.1-mini, introducing native audio reasoning and reducing p95 latency by 25%. Developers can configure reasoning effort to control latency and token usage.
Impact: High
Why it matters
You can now build low-latency voice agents that reason internally and explain actions before executing tools, reducing user interruptions.
TL;DR
- 01Add conversational preambles to avoid silence when executing tools.
- 02Lower p95 latency by 25% using improved context caching.
- 03Opt for gpt-realtime-2.1-mini to cut audio output costs by 3x compared to the full 2.1 model.
Key facts
- gpt-realtime-2.1-mini audio output rate
- $20.00 per 1M tokens
- gpt-realtime-2.1 audio output rate
- $64.00 per 1M tokens
- p95 latency reduction
- at least 25%
- gpt-realtime-2.1-mini cached audio input
- $0.30 per 1M tokens
- gpt-realtime-2.1-mini fresh audio input
- $10.00 per 1M tokens
Realtime 2.1 Mini Reasoning Integration
The gpt-realtime-2.1-mini model is a low-latency reasoning model designed for direct voice-to-voice interfaces. It processes audio and text inside a single model pipeline, avoiding traditional multi-model latency (speech-to-text combined with text-to-speech text generation). The model is capable of tool calling (function calling) with configurable reasoning efforts, enabling multi-step actions.
P95 Latency and Input Caching
OpenAI has implemented improved prompt caching that reduces p95 latency by at least 25% across all Realtime voice models. Caching also minimizes operational costs in longer sessions since the system prompt is cached after the first turn. Audio input tokens that are cached cost $0.30 per 1M, while fresh un-cached audio input costs $10.00 per 1M. Audio output for the mini model is priced at $20.00 per 1M tokens.
WebRTC Peer Connection Setup
Developers can initiate WebRTC audio streams directly from browser clients. To avoid exposing production API keys, the backend server must first mint a short-lived ephemeral token using the /v1/realtime/client_secrets endpoint. This token is then passed to the frontend to establish a WebRTC PeerConnection.
Try it in 2 minutes
const r = await fetch("https://api.openai.com/v1/realtime/client_secrets", {
method: "POST",
headers: {
"Authorization": `Bearer ${process.env.OPENAI_API_KEY}`,
"Content-Type": "application/json"
},
body: JSON.stringify({
model: "gpt-realtime-2.1-mini",
instructions: "You are a support agent. Reply in one or two short sentences."
})
});
const { value: EPHEMERAL_KEY } = await r.json();javascript
✓ When to use
- Building conversational voice assistants requiring multi-step tool calls.
- Implementing low-latency WebRTC speech applications on mobile or web.
✕ When NOT to use
- Standard text-only agents where simpler chat completion endpoints are more cost-effective.
- Scenarios with highly strict budget constraints where audio APIs are still too expensive.
What to do today
- Migrate existing realtime-mini setups to gpt-realtime-2.1-mini.
- Configure reasoning_effort to 'low' as a default for speech-to-speech pipelines.
- Implement token caching optimizations to leverage the $0.30/1M pricing tier.
Sources