Phonic vs Grok Voice vs Gemini Live: which voice model sounds most like a real call?
When your agent calls you, a realtime voice model does the talking. We gave three of them the same four prompts and timed how long each one took to start speaking, with and without tools.
Why the model matters on a call
agentring rings your phone like a real call. When you answer, a voice AI explains why your agent is calling, answers your questions and, when it needs to, asks your agent or one of its tools. Two things decide whether that feels like a conversation or like waiting on hold:
- How fast it starts talking after you stop. Past about a second and a half, the pause gets noticeable.
- Whether it uses tools well: picking the right one, chaining two when needed, and not going quiet while a slow one runs.
agentring can use Gemini Live, OpenAI Realtime, Grok Voice or Phonic, so we compared them on exactly these two points.
How we measured
On October 30, 2026 we streamed the same four Spanish prompts, in real time, into Phonic-1, Grok Voice and Gemini 3.8 Live, three sessions each:
- plain conversation, no tools;
- a question that needs one quick tool;
- a question that needs two tools, one after the other;
- a question whose tool takes 3 seconds to answer.
First audio is the time from the end of the user's speech to the first audio from the model. Right tools counts the tool turns (three prompts × three sessions = 9) where the model picked and chained the tools we expected. OpenAI Realtime was not part of this run.
Results
Time to first audio, in seconds (lower is better)
- Chat, no tools
- With tools (range over the tool prompts)
| Phonic-1 | Grok Voice | Gemini 3.8 Live | |
|---|---|---|---|
| First audio, chat | 1.18 s | 1.26 s | 1.60 s |
| First audio, with tools | 1.20–1.52 s | 1.30–1.59 s | 2.00–5.30 s |
| Right tools | 8/9 | 9/9 | 7/9 |
What it sounds like
- Phonic and Grok fill the wait. Before calling a tool they say something like "let me check", then answer when the tool returns. You never sit in silence.
- Gemini goes quiet until the tool answers. That's most of its 2–5 s with tools: with the 3-second tool you hear nothing until the result is in.
- Phonic talks the longest, Grok is the most concise. On a call where you just want the decision ("roll back or not?"), shorter is usually better.
- Grok picked the right tools every time (9/9), Phonic missed once and Gemini twice.
What it costs
- Gemini Live has a free tier (rate-limited; Google may use free-tier data to improve its products) and a paid tier of about $0.005 per minute of audio in and $0.018 per minute out (Google's pricing).
- Phonic starts at $0.07 per minute (phonic.ai).
- Grok Voice and OpenAI Realtime bill per use; see xAI's and OpenAI's pricing pages.
Which one should you pick?
- Trying agentring out on your own server: Gemini Live. The free tier costs nothing to start with, and it's the default.
- Calls that lean on tools (asking your agent, checking CI, looking things up): Grok Voice. Fast, concise and the most accurate with tools in this run.
- The quickest first word: Phonic, by a small margin, if you don't mind longer answers and the higher price.
In agentring it's one setting, the same for the CLI and the MCP server: --provider gemini|openai|grok|phonic, or AGENTRING_VOICE_PROVIDER in the environment. With agentring cloud the voice AI is included in the plan, so there's no key to set up.
Caveats
This is a small test: four prompts, three sessions per model, in Spanish, from one machine on one day. Realtime models change often and latency depends on the network and the region you reach. Treat the numbers as a snapshot and measure with your own prompts if the choice matters to you. The provider setup and the latest notes are in the voice docs.
Let your agent call you. Free to start, open source underneath.