What does an hour
of conversation cost?
The gap between five cents and half a cent per minute is large. So is the gap between the services those prices buy.
OpenAI lists GPT-Live-1 at $0.05 per minute of voice-session time, billed per second. Separate backend reasoning and tool usage cost extra. An hour of connected voice time is $3.00 before those charges.
.wave lists an introductory $0.005 per minute for Nemotron VoiceChat 11B. That is $0.30 per hour, or one-tenth the session price. In its current public beta, sessions last at most 120 seconds. The API does not support tools, custom instructions, or text input.
| Service | Per minute | Per hour | 100,000 minutes |
|---|---|---|---|
| GPT-Live-1 | $0.050 | $3.00 | $5,000 |
| .wave Nemotron beta | $0.005 | $0.30 | $500 |
| Duplex Compute target | $0.010 | $0.60 | $1,000 |
Duplex Compute is in development. Our target price is a hypothesis we are testing with potential customers, not an available offer. The comparison does not imply equal capabilities, quality, or reliability.
Tools need a path back into the conversation
The underlying open NemotronLabs model supports a function channel. That does not make every hosted version tool-capable. A service must expose the calls, accept results, and preserve the conversation while a tool runs. For an application developer, that integration is part of what the API needs to provide.
Structured calls are only the beginning. The system must handle a corrected date, a slow search, a cancelled request, and a failed payment without speaking as though an action succeeded. Reliability on these cases can matter more than the headline per-minute price.
GPU price is not conversation price
A GPU rented for $2 per hour costs $0.50 per conversation-hour when it serves four simultaneous conversations continuously. If those four slots average 25% utilization, the same machine costs $2 per conversation-hour. These are illustrative calculations, not throughput claims about our service.
A warm $2/hour machine costs about $1,460 over a 730-hour month. At a proposed $0.01 per minute, covering that GPU alone requires about 146,000 billable minutes, or 2,433 conversation-hours. That averages 3.33 concurrent conversations across the month. Payment fees, storage, engineering, and support are still additional.
Per-second rentals make short experiments inexpensive. They do not remove the time needed to load models or guarantee a warm machine when someone starts a call. Efficient serving and predictable demand determine the cost of an API people can rely on.
The gap we’re exploring
We’re interested in lower-cost voice that supports application-defined tools, configurable behavior, and longer conversations. We’re testing demand before committing to a particular model or serving stack.
For learning apps, support tools, and conversational products, the right question is how much a successfully completed interaction costs. A cheap minute that loses the user’s request can cost more than an expensive one that gets it right.
Request early access and tell us your expected volume, use case, and current limitations.