All capabilities Image, audio and video

Talk to it out loud and hear it answer

Not transcribe-then-answer-then-synthesise. One session on gpt-realtime-2.1 where audio goes in and audio comes back, with the turn-taking handled for you.

DocsOfficially documentedTestNot tested by benchr Deep · API Docs checked September 1, 2026
OpenAI API

Overview

  • Runs over three transports: WebRTC for browser and mobile clients, WebSocket for server-side pipelines, and SIP for telephony.
  • Removes the stitched pipeline, which is what removes most of the latency people notice in voice agents.
  • Exposes reasoning effort as a knob, so a production voice agent can be tuned for responsiveness.

How it works

  1. 01Pick the transport by where the audio lives - browser to WebRTC, server to WebSocket, phone number to SIP.
  2. 02Start with reasoning effort set to low, which the documentation names as the production starting point for voice agents.
  3. 03Tune the transcription delay setting deliberately: lower delay gives earlier partial text, higher delay improves transcript quality.
  4. 04Design the interruption behaviour before the prompt. Barge-in is the difference between a demo and something people will use.

Limits

  • It is a session protocol, not a request-response endpoint - reconnection, state and audio buffering are your responsibility.
  • The latency and quality trade-off is explicit in the documentation and cannot be optimised away, only chosen.
  • Telephony via SIP puts you inside carrier constraints the model documentation does not cover.

Verification

benchr test runs

benchr has not run this capability. No test has been published.

benchr test runs →

Evidence

Not stated by the source

  • End-to-end latency figures for any transport
  • Concurrency limits per organisation
Source-check history
  • September 1, 2026DocsOfficially documentedFirst logged. Model id, the three transports and the low-reasoning-effort production guidance read from the realtime guide.