use cases · phone reception, right stack ·
A company selling an AI receptionist to practices pays for every minute its agent is on the phone: about €7,000 a month at 100,000 minutes with OpenAI gpt-realtime-2.1. We put six voice stacks through the same 50 French calls with a simulated patient, 59 lines recorded by two of us, and live calls from our own phones, and priced every minute.
63% less a minute
Deepgram, GPT-6 Luna and Cartesia, assembled: 46 of 50 calls handled well, no serious error, and the fewest silent mistakes on our own voices. €0.026 a minute against €0.070.
the catch
It is slower to answer, and runs in the US without HDS. Gemini 3.8 Live has both HDS and the EU and costs 38% less than the references, but made 2 serious errors in 50.
an AI receptionist for practices · 100,000 minutes a month
new bill−63%
€0.026 a minute instead of €0.070
The catch: slower to answer, and hosted in the US without HDS.
In the three-part stack the language model is 2% of the bill: the voice is the cost. Speech-to-speech models bill the whole call again at every turn. Every option bills by use, so the gap stays the same at any volume.
four finalists
Deepgram, GPT-6 Luna, Cartesia three parts, assembled
€0.026
Gemini 3.8 Live speech to speech
€0.043
OpenAI gpt-realtime-2.1 speech to speech
€0.070
ElevenLabs Agents GPT-6 Luna inside
€0.072
| option | handled well | serious | wrong on our voices | wait | € a minute | at 100k minutes | health data |
|---|---|---|---|---|---|---|---|
| Deepgram, GPT-6 Luna, Cartesia three parts, assembled | 46 of 50 | 0 | 4 of 59 | 2.4 s | €0.026 | €2,600 | US, no HDS |
| Gemini 3.8 Live speech to speech, Google Cloud | 42 of 50 | 2 | 5 of 59 | 1.2 s | €0.043 | €4,340 | EU, HDS |
| OpenAI gpt-realtime-2.1 speech to speech | 45 of 50 | 0 | 6 of 59 | 1.7 s | €0.070 | €7,040 | US, no HDS |
| ElevenLabs Agents GPT-6 Luna inside | 48 of 50 | 0 | 5 of 59 | 2.4 s | €0.072 | €7,170 | US, no HDS |
slower
2.4 s at the median before it answers. Gemini Live: 1.2 s; OpenAI realtime: 1.7 s.
health data
It runs in the US, without HDS. Gemini 3.8 Live has HDS and the EU, but made 2 serious errors in 50.
a small test
On 50 calls, two options are known to about ±10 points: these results rule out a large drop, not any drop.
cheaper is not enough
Two cheaper stacks, OpenAI gpt-realtime-mini and Voxtral with Mistral Small 4, handled only 25 and 24 of 50 calls well.
Handled well: no error, or one a front desk would shrug off. Graded against a written rubric and the booking system's final state, checked by hand on 15 calls.
Wrong on our voices: a surname, date of birth or phone number captured wrong without asking again, on 59 lines recorded by two of us. The three-part stack asked again on 15 of 59 instead.
Wait: the median, from the end of the caller's speech.
Blind listen, two of us on 10 clips a voice (naturalness out of 5; share fit for patients): ElevenLabs Agents 4.7, 100%; the three-part stack 3.8, 100%; Gemini 2.1, 45%; OpenAI realtime 1.9, 50%. 12 live calls by two of us on Gemini and the three-part stack: none broke, and one on the three-part stack went wrong.
Synthetic data: a simulated patient and a fictional practice. No client data.
Full study on request: hugo@jayn.app
*on your own data and at your volume. you only pay her if she generates savings.
the cheapest option
that meets your quality bar, and its bill.
what to fix first
before switching, so it does not cost you customers.
your own model
whether it pays, and from what volume.