Session time and audio time
have different prices.
A per-minute voice price can meter an open connection, incoming audio, or generated speech. The hourly numbers in our table keep those billing rules visible.
| API / model | Rate | Hourly figure in the comparison |
|---|---|---|
| GPT-Live-1 | $0.05/min of connected voice-session time | $3.00; backend reasoning and tool usage extra |
| Gemini 3.1 Flash Live Preview | $0.005/min audio input + $0.018/min audio output | ≈$0.84 in audio charges at 60 min input + 30 min output |
| Gemini 3.8 Live | $0.005/min audio input + $0.018/min audio output | ≈$0.84 with the same audio assumptions |
| Gemini 3.8 Live Extended Thinking | Same audio rates; text output including thinking is $4.50/M tokens | ≈$0.84 in audio charges; text/thinking usage extra |
| Grok Voice Think Fast 2.0 | $0.08/min + $0.004/text input message | $4.80 with automatic turn detection, before text and tool charges |
| .wave Nemotron VoiceChat 11B | $0.005/min while the session is open; introductory beta rate | $0.30 equivalent; each session lasts at most 120 seconds |
| .wave Nemotron 3.5 ASR Streaming 0.6B | $0.00045/min from first audio until session closes | $0.027 for transcription only; no reasoning or speech output |
Gemini: count both audio streams
Google lists the same audio rates for these three Live models: $3/M input audio tokens, equivalent to $0.005/min, and $12/M output audio tokens, equivalent to $0.018/min. Text input costs $0.75/M tokens. Text output, including thinking tokens, costs $4.50/M tokens.
Our example assumes an hour of continuously streamed input and 30 minutes of generated speech: 60 × $0.005 + 30 × $0.018 = $0.84. At 15 minutes of speech output, the audio subtotal is $0.57. At 45 minutes, it is $1.11. Streaming less input changes the input charge too.
These figures exclude text and context usage, thinking, video, and search. They are workload illustrations, not measured Understand invoices or a fixed hourly offer. The Extended Thinking row shares the audio rate, not necessarily the final bill.
Grok: speed control and session billing
The speed-adjustable voice API researched for Understand is Grok, from xAI. Its speech-to-speech guide identifies grok-voice-think-fast-2.0 as the flagship model and documents audio.output.speed from 0.7 to 1.5, with 1.0 as the default.
Automatic turn detection bills session duration. Push-to-talk bills only audio sent and received. The published audio rate is $0.08/min, or $4.80 for an hour under the session-duration rule. Text input costs $0.004/message; function-call results are exempt. Search and other tool services can add charges.
Two .wave models, two different services
VoiceChat 11B listens and speaks in full duplex. Its $0.005/min rate is introductory. Sessions last at most two minutes, and tools, custom instructions, and text input are unavailable. The $0.30 hourly equivalent does not imply that an hour-long session is supported.
Nemotron 3.5 ASR Streaming 0.6B transcribes audio in 32 languages. At $0.00045/min, an hour costs $0.027. A conversational pipeline still needs a reasoning model and speech synthesis, so this rate belongs in the component table rather than the complete-agent comparison.
Prices that providers haven’t published
Thinking Machines describes TML-Interaction-Small as a research preview without a public API rate. We label that row “Price not published.” Its score does not supply a serving price.
Self-hosted Nemotron has no fixed per-minute API bill. Divide the GPU hourly rate by average occupied concurrent streams to estimate GPU cost per conversation-hour. For example, a hypothetical $1/hour GPU averaging four occupied streams costs $0.25 per conversation-hour before hosting overhead. That is arithmetic, not a tested Nemotron capacity or a rental quote.
Gemini 3.8 and Grok have no matched FDB-v3 results in this table. The Understand research used other benchmarks; their percentages cannot be inserted into the FDB-v3 columns.