The OpenAI Realtime API builds speech-to-speech voice agents with barge-in (interruption), low first-audio latency, automatic turn detection, and realtime tool use, over WebRTC or WebSocket. The current model is gpt-realtime-2.1, priced at $32/1M input and $64/1M output audio tokens (a mini version at $10/$20). Pick it when you need a managed, conversational voice agent and don't want to build turn-taking and interruption handling yourself.
Other good fits
A voice agent that needs to call tools mid-conversation, with tool execution kept on your own backend.
+ Pros
Read pros
The Realtime API is OpenAI's platform for speech-to-speech voice agents: it handles barge-in (letting a user interrupt mid-response), low first-audio latency, natural turn-taking through voice activity detection, and tool calls executed directly within an audio conversation. Sessions connect over WebRTC or WebSocket, following a flow where your server creates an ephemeral client secret, your frontend initializes a session, and the connection is established from there, keeping tool execution and session management on your own backend rather than exposed to the client. The current model is gpt-realtime-2.1, which maintains conversation state across a session.
Pick it when…
You need natural interruption handling (barge-in) and low first-audio latency without building that logic yourself.
Your voice agent needs to call tools mid-conversation while keeping execution server-side.
You want automatic turn detection instead of building your own voice-activity detection.
− Cons
Read cons
This is a hosted, proprietary API: you can't self-host it or inspect how it works internally, and you depend on OpenAI's availability and pricing. Per-token audio billing means a chatty or long-running conversation costs meaningfully more than a short, focused one, so unbounded conversations are a real cost risk without a length or turn cap.
Pick something else when…
You want a self-hosted, open alternative.Moshi is a genuinely open, self-hosted full-duplex voice model, at the cost of your own GPU and setup effort.
You want a framework to build a custom voice pipeline around any model, not just OpenAI's.Pipecat is built for assembling realtime voice pipelines from various components.
You want a different managed provider's realtime API to compare on latency and price. Compare against another hosted conversational voice API before committing.
Avoid if
Your conversations run long; usage charges accumulate with conversation duration, so estimate cost before building around it.
Context
gpt-realtime-2.1 is the current model; a lower-cost gpt-realtime-2.1-mini variant is also available. For applications with identified end users, OpenAI recommends sending an OpenAI-Safety-Identifier header to improve safety enforcement targeting.
Requirements
Sessions need a server component (to mint the ephemeral client secret and keep tool execution off the client) plus a frontend that initializes the RealtimeSession over WebRTC or WebSocket. This is a hosted API, not something you self-host; you need an OpenAI API key and network access, not local GPU hardware.
Billing is per audio token, not a flat per-minute rate, so cost scales with how much is actually said on both sides of the conversation. Long conversations accumulate cost quickly; estimate expected conversation length and volume before committing to a design, and consider the mini model for less demanding use cases.
First task
Documentation-based suggestion — not run here.
Build a short, scoped voice conversation with one tool.
Set up the ephemeral-secret flow on your server and connect a test frontend session.
Give the agent one narrow tool, for example looking up a fixed, fictional fact.
Have a short test conversation, including one deliberate interruption mid-response.
Check that the interruption was handled naturally and the tool call only fired when actually requested.
Done when the interruption test behaves naturally (the agent stops and responds to the new input) and the tool was called only when appropriate, verified against the conversation transcript.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on OpenAI's Realtime API guide and pricing documentation (developers.openai.com, redirected from platform.openai.com/docs), checked 23 September 2026. No session was created and no conversation was run for this review.
Moshi wins when self-hosting and full model access matter more than a managed, polished API.
Pipecat wins when you want a flexible framework to assemble a voice pipeline from multiple components, potentially including OpenAI Realtime as one piece.
Gemini Live wins when you want to compare Google's realtime voice offering against OpenAI's on price and features.
Stream microphone audio and receive spoken responses in one API.
Build interruptible voice conversations with function calling.
Let agents look up information or trigger actions while speaking.