Short answer: all four Qwen Audio realtime models heard the same spoken question correctly and began answering about half a second after it ended. The two Flash models cost under a tenth of a cent for the turn, and the two Plus models under a cent. Qwen Audio 3.1 Realtime Flash, Qwen Audio 3.1 Realtime Plus, Qwen Audio 3.0 Realtime Plus and Qwen Audio 3.0 Realtime Flash all run on EmpirioLabs through the same realtime endpoint. Listen to each reply below; we are not declaring a winner.
The question
"I have a nine a.m. flight and the airport is forty minutes away. When should I leave home?"
Listen to the four replies
"You should leave home by 7:20 AM to account for the 40-minute drive and a standard two-hour arrival buffer. This ensures you arrive at the airport by 8:00 AM, well before your 9:00 AM flight."
"You should leave home by 7:40 AM to arrive at the airport by 8:20 AM. This gives you a 40-minute buffer to check in and go through security before your 9:00 AM flight."
"You should leave home at least 2 hours before your flight to account for check-in and security. Considering the 40-minute drive, aim to depart around 6:20 AM for a 9:00 AM flight."
"You should leave home by 7:40 AM to account for the 40-minute drive and typical airport check-in time. This gives you a buffer to arrive at least an hour before your 9:00 AM flight."
Results at a glance
| 3.1 Realtime Flash | 3.1 Realtime Plus | 3.0 Realtime Plus | 3.0 Realtime Flash | |
|---|---|---|---|---|
| Price for this turn | $0.0008 | $0.0080 | $0.0085 | $0.0007 |
| First audio after the question ended | 0.48 seconds | 0.54 seconds | 0.49 seconds | 0.52 seconds |
| Length of the spoken reply | 13.1 seconds | 10.5 seconds | 11.4 seconds | 10.6 seconds |
| Voices | 88 | 88 | 19 | 19 |
| Web search with function calling in one session | Yes | No, one or the other | No, one or the other | No, one or the other |
| Input and output language settings | Yes | Yes | No | No |
How we ran it
The question was spoken once (synthesized with Qwen Audio 3.1 TTS) and streamed to each model in real time as 16 kHz PCM, followed by a second and a half of silence, over wss://api.empiriolabs.ai/v1/realtime. Each session used the same instructions ("You are a helpful voice assistant. Answer in English in two short sentences."), text and audio output, and server turn detection with a 900 millisecond pause. One shot each, on November 2, 2026. The first-audio time runs from the model detecting the end of the question to the first audio of the reply, measured at the client. Prices are the amounts charged for these exact turns.
What to look for
The arithmetic. Leaving at 7:40 with a forty-minute drive arrives at 8:20, forty minutes before the flight; leaving at 7:20 arrives at 8:00, one hour before. Compare each reply's plan with the buffer it says it gives.
Flash against Plus. Within each generation, Flash is priced well below Plus. In this test the four started speaking within a few hundredths of a second of each other.
Run the same test on EmpirioLabs
wss://api.empiriolabs.ai/v1/realtime?model=qwen-audio-3-1-realtime-flash
Authorization: Bearer YOUR_EMPIRIOLABS_API_KEY
{"type": "session.update", "session": {
"voice": "xunanchuan_v3.1",
"instructions": "You are a helpful voice assistant. Answer in English in two short sentences.",
"turn_detection": {"type": "server_vad", "silence_duration_ms": 900}
}}Then stream microphone audio as input_audio_buffer.append events. Swap the model in the URL to try the others, using a 3.0 voice such as daniel on the 3.0 models, or click Start session in the playground.
Frequently asked questions
Were the replies edited or retried?
No. Each reply is the model's first answer to the question, unedited.
Why do the Plus models cost about ten times as much?
Their audio and text tokens are priced higher. The token counts for this turn were similar on all four.
Which one should I use?
Try the question your users actually ask. The launch guides cover each model's settings: Qwen Audio 3.1 Realtime Flash and Qwen Audio 3.0 Realtime Plus and Realtime Flash.
