Talk to it. It talks back.
A thread's agent can be a realtime voice model. What it heard and what it said are lines in the thread, and the session's audio seconds and cost are its last record.
One session
One engine, any realtime model.
The harness declares the model. The host opens the provider's realtime socket with your own key, and the session engine turns its events into the thread's records.
What it heard, what it said.
A live session with OpenAI's gpt-realtime-2 on this computer, run by the live test. The operator's name is left out.
22:23:53.580 note person:operator: Hello. 22:23:53.812 note person:operator: Can you hear me? 22:23:55.154 note person:operator: What is 2 plus 2? 22:23:55.995 note agent:openai-gpt-realtime-2@livebox: Yes, I can hear you. Two plus two is four.
Case study · two providers, one engine
The same question, two voices.
The same live test against two providers' realtime models, one day apart. Both heard the question and answered it. The cost a minute is the session's own usage record.
Talk over it. When you speak over the agent, the engine records an interrupted event and the agent stops. These two sessions did not test it, so no number for it is shown here.
- OpenAI
gpt-realtime-2: 6.45 s of audio in, 3.75 s out, $0.0091 - 4.3 s
- xAI
grok-voice-3.0: 6.05 s of audio in, 3.51 s out, $0.08 - 3.7 s
- a minute on OpenAI, $1.29 a minute on xAI (xAI bills started minutes)
- $0.13
test-results/realtime-host/ (this computer)Output case-studies/realtime/stats.jsonTool the live realtime test, read by scripts/sites/case-studies.mjs realtimeRan on This computer (dell-gb10-1), the providers' realtime APIsWhen Talk to your agent.
Pick a realtime agent in a thread and press the microphone.