Fluid
Reliable voice-to-voice.
It knows when to talk.
- CallerHi, I need to move my appointment on Thursday, something came up at work.
- FluidListensMhm.
- FluidStopsNo problem. Would Friday morning work, or rather next week?
- CallerNext week, ideally Tuesday afternoon.
- FluidWord for wordYou’re booked for Tuesday at 3 pm. Please cancel at least 24 hours ahead, otherwise the appointment is charged.
- CallerGreat, thank you.
Hear a real call.
One take from our playground, unedited: the weather, plans for the day and a phone number read back.
Frontier-level on τ²-bench.
Fluid by voice scores within five points of the same frontier model in chat.
- All tasks
- 84%89%
- Airline
- 86%86%
- Retail
- 76%82%
- Telecom
- 90%98%
τ²-bench v2, English, 150 tasks, one attempt each, spoken with tau2’s simulated customer in its two accented caller personas. Frontier model: Claude Opus 5.5 in both. Measured in November 2026.
Next to other voice agents.
Fluid 1 on τ²-bench’s 150 English voice tasks, beside every voice agent on its official leaderboard.
- Ava ZeroAvartha
- Phonic 1Phonic
- GPT Live 1OpenAI
- Fluid 1KugelAudio
- Gemini 3.8 LiveGoogle
- Grok Voice Think Fast 2.0xAI
Other voice agents: their English results on the official τ²-bench voice-v2 leaderboard as of November 2026, with the same tasks and simulated customer. Fluid’s run is our own measurement, not a leaderboard submission.
Built for real calls.
Backchannels
A short mhm while the caller is talking.
Follows the language
Fluid switches between German and English when the caller does.
Transcripts
Both sides of the call arrive as text.
OpenAI Live compatible.
client = AsyncOpenAI( api_key=os.environ["KUGELAUDIO_API_KEY"], base_url="https://api.kugelaudio.com/v1", ) async with client.live.connect() as connection: await connection.session.start(session={ "model": "fluid-1", "instructions": instructions, "audio": {"output": {"voice": "973"}}, })
Pricing
€0.05per minute
€0.0595 including 19% German VAT
Back-office model costs are not included.
Need volume pricing or on-premise?
Questions
What is Fluid 1?
Fluid 1 is KugelAudio’s speech-to-speech model for voice agents. It hears your customer, answers in a natural voice and hands what it can’t answer to your backend.
Can I use my own model and tools?
Yes. Your backend answers with any model you choose and runs your tools. Or let us run the model, and keep only the tools on your side.
How do I move from OpenAI Live?
Fluid follows the OpenAI Live protocol. Point the official OpenAI SDK at Kugel’s address, use your Kugel key and set the model to fluid-1.
Does it work on the phone?
Yes. Fluid takes phone audio at 8 kHz as well as 16 and 24 kHz audio from apps and browsers, and answers in the same format.
How is it billed?
€0.05 per minute of conversation, billed by the second, from the first audio until the session ends. A model you run in your own backend is billed by its provider.