Voice-controlled assistants were long a topic for corporations with their own research budgets. That is currently changing on the price side: Google Deepmind has presented two audio models for developers with Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. According to the analysis firm Artificial Analysis, the variant with Extended Thinking leads the speech-to-speech leaderboard — that is, the ranking of those models that translate spoken input directly into spoken responses without the detour via a separate text intermediate step.
In terms of price, Google states 1.38 dollars per hour of voice conversation. Calculated, that is a good two cents per minute of conversation. This places the offering well below GPT-Live-1 from OpenAI. Its strength lies elsewhere: full-duplex operation, in which model and human can speak simultaneously and interrupt each other, is likely to enable more natural conversations.
Why the price changes the discussion
A phone call with a support request rarely lasts more than a few minutes. With costs in the low cent range per minute, the question shifts away from "Can we afford this?" to "Where does it actually bring benefits?". That is a different type of project: no longer a flagship initiative with its own business case, but a building block that can be integrated into existing processes — appointment scheduling, status information on an order, pre-qualification of enquiries outside office hours.
The distinction between model costs and total costs remains important. The minute of conversation is only one item. Added to this are telephony integration, hosting, monitoring, maintaining the knowledge base and — mostly underestimated — the effort for testing and fine-tuning. An inexpensive model does not automatically make a project inexpensive, but it lowers the threshold for even getting started.
Latency is the actual quality characteristic
In text chats, one second of delay is forgiven. On the phone, it is not. People interpret pauses as uncertainty, disinterest or as a signal to keep talking themselves. Speech-to-speech architectures address exactly this: anyone who first converts speech into text, sends this through a language model and then voices the result again collects delay at every station. The direct path from audio to audio saves these stages.
Full-duplex capability belongs in the same category. A conversation in which you cannot interrupt the other party feels like a voice menu with a better voice. Anyone planning with voice agents should therefore not select models solely by benchmark position, but test them in their own use case: with real background noise, with dialect, with people who change their minds mid-sentence.
Conversation design is conceptual work, not a model question
A good model does not conduct a good conversation. It only provides the language capability. What is said, when follow-up questions are asked and when a handover to a human takes place is conceptual work. A few points have proven viable in this regard:
- Clear scope of responsibility: An agent that only books appointments, but does so reliably, is more valuable than one that can do everything halfway.
- Defined exits: Rules are needed for when the conversation is handed over to staff — in the case of complaints, repeated misunderstanding, at the request of the calling person.
- Confirmation steps for binding actions: Before an appointment is firmly booked or an order is changed, the agent should summarise and ask for confirmation.
- Transparency: Callers should know that they are speaking with a system. This is not only a legal question but also a question of expectations.
Integration into existing web and CMS systems
The technically most exciting part is rarely the language model, but the interfaces. A voice agent is only as good as the data it is allowed to access. For information on opening hours, products or services, this is often the existing content in the CMS — in TYPO3, for example, structured via pages, content elements and custom records.
This gives rise to a requirement that improves many projects anyway: content must be machine-retrievable and unambiguous. Where opening hours appear in three variants on three pages, an agent will reliably find the wrong one. A clean source of truth, sensibly modelled records and an API that reflects the editorial status are the actual preparatory work.
For actions — creating an appointment, querying order status, noting a callback — defined function calls against the specialist systems are needed, with permission checks and logging. And an answer is needed to the question of what happens when the target system is currently unreachable. An agent that politely promises a callback in this case is better than one that invents a booking.
Data protection and logging
Voice data is personal and often significantly more sensitive than form entries. Before launch, three questions need to be clarified: where audio data is processed, how long recordings and transcripts are retained, and who in the company may view them. For quality assurance, conversation logs are needed — for data protection, deletion periods are needed. Both can be reconciled, but not improvised after the fact.
A realistic entry point
A narrowly scoped pilot makes sense: one use case, one target group, one measurable criterion. For example, appointment scheduling outside service hours, measured by the number of successfully completed bookings and the handover rate to staff.
From these figures it is possible to derive seriously whether an expansion is worthwhile — and which conversation situations the model does not yet master.
For web and software projects, the development means above all one thing: language becomes another channel on the same content and specialist data, not a separate product.Anyone who structures their content cleanly in the CMS and makes their specialist systems accessible via robust interfaces can add voice agents later without rebuilding.And because models and prices are visibly moving quickly, the connection should remain interchangeable — the investment lies in data quality and conversation design, not in the model of the respective quarter.