Which Speech-to-Speech Model Should You Choose for Long Conversations?

Benchmark of October 2026 · GPT-Live-1, Grok Voice 3.0, Gemini 3.8 Live and Gemini 3.1 Flash Live · Calls in French

• 460+ minutesof calls on four models, read question by question
• 32 pathsof questions in 33 calls, with the same buyer and the same prompt
• 4 rulesof prompt writing that make a long questionnaire more reliable

The starting point: a voice to choose, a script to hold

A property search assistant qualifies buyers by phone. In a call of 6 to 12 minutes, it has 24 questions to ask about the buyer's project. Every question it does not ask is a piece of information missing at the end of the call.

For this kind of call, speech-to-speech gives the most natural experience. One model hears and answers in voice, with no transcription and no synthetic voice added at the end: the tone follows the conversation, the model handles interruptions itself, and the answer comes in one step instead of three (see our page on speech-to-speech AI voice agents).

Choosing a speech-to-speech model then comes down to two criteria. The first is how it sounds: its voice, its rhythm, the way it reacts. You and your callers judge it by listening; this benchmark does not measure it. The second is reliability. These models are less deterministic: the same call, replayed, does not take the same path, and on a long script some questions may never be asked.

This article measures the second criterion on four models: GPT-Live-1, Grok Voice 3.0, Gemini 3.8 Live and Gemini 3.1 Flash Live. Its question is practical: if you prefer a model for its voice, how far can the prompt make it reliable, and when should you change model? We measured in two stages: calls with the original prompt, then the same calls once the fragile turns of the script were rewritten.

The method

The agent. The script has 24 questions, 27 with their follow-ups, and the prompt about 21,000 characters. The prompt is the written instructions the agent follows. It is the kind of agent our page on pre-qualifying calls describes.

The calls. The same agent, with the same prompt, runs on the four models, launched in waves of one call per model. Gemini 3.8 Live is the default Google recommends for low-latency voice agents (Live API skill, migration guide). All four call the same simulated buyer, itself a model, which answers from a fixed sheet; the agent gets the same context about the buyer every time.

Two series. 33 calls with the original prompt, 8 or 9 per model, then 24 with the adjusted prompt, 6 per model: more than 460 minutes of conversation.

First finding: from one model to another, 89% to 98% of questions asked with the original prompt

With the original prompt, the four models ask between 89% and 98% of the due questions. What varies from one call to the next is the path: same buyer, same prompt, the 33 calls follow 32 different paths.

Original prompt, 8 or 9 calls per model:

• GPT-Live-1 asks the most and has 99% of the information at the end of the call. It skips significantly fewer questions than Gemini 3.1 Flash Live (-4.38 to -0.12 per call), the only significant gap of this series. It is also the one that makes the caller repeat the most.
• Grok Voice 3.0 makes the caller repeat the least and is the most regular: 19 questions out of 27 have the same fate in all its calls.
• Gemini 3.8 Live asks as many questions as Grok Voice 3.0, but makes the caller repeat 1.44 times per call, against 0.38.
• Gemini 3.1 Flash Live is the least regular: 6 questions out of 27 with the same fate in all its calls, 12 to 21 questions asked depending on the call. A single one of its 8 calls skips 10.

The same call, replayed, does not take the same path: judge a model on several calls, never on one.

What the prompt can fix: four fragile turns, four rules

The reviewers sorted every lost question by what caused it. Part of the losses come from the way the script is written, in four kinds of turns: a follow-up planned “on the next turn”, a counter-question from the buyer never looped back, a question replaced by the announcement of a task, and a block of questions with no transition sentence. Figure 3 counts the losses of the first three by model; they do not trip the same models:

• Gemini 3.8 Live is the most exposed, above all to counter-questions (7) and to the task (5): it asks the same late question without getting the answer in 7 calls out of 9.
• Grok Voice 3.0 stumbles mostly on the task: 4 questions replaced by its announcement. The question before it, on what weighs most on the buyer, is skipped in 4 calls out of 8.
• Gemini 3.1 Flash Live forgets only one follow-up. Its losses come mostly from one call where it cuts the path short by 10 questions.
• GPT-Live-1 has only 3 failures in 8 calls.

These turns are written in the script, so that is where they are fixed. Five passages of these four kinds are rewritten, and the rest of the prompt does not change. Each turn has its rule:

Second finding: from one model to another, not the same ability to improve with the prompt

On the five rewritten passages, all models together, the information is obtained in 82% of cases before (136 out of 165), 91% after (109 out of 120). Across all questions, the share of due questions asked stays at 94% (93.8% before, 94.4% after). Re-asks go from 1.33 to 1.00 per call, skipped questions from 1.24 to 0.79.

In numbers, on the three measures:

Model by model, what the adjustment changes, and what remains:

• Gemini 3.1 Flash Live gains the most: from 89% to 96% of questions asked, and from 33 out of 40 to 29 out of 30 on the rewritten passages. It still skips the question on the calendar twice.
• GPT-Live-1 makes the caller repeat significantly less: from 2.25 to 1.33 re-asks per call (-1.71 to -0.17). It still asks almost every question, and still makes the caller repeat more than Grok Voice 3.0 and Gemini 3.1 Flash Live (0.33 and 0.50).
• Gemini 3.8 Live progresses on the rewritten passages (31 out of 45, then 25 out of 30), not overall: 94% then 93% of questions asked. It still skips the question on what weighs most on the buyer in 3 calls out of 6, and once a whole block.
• Grok Voice 3.0 still makes the caller repeat the least, 0.33 times per call, and skips only 4 questions in 6 calls. But it ends one call itself mid-script, leaving 8 questions unanswered: hence 94% then 90%.

The prompt makes a model more reliable up to a point: it fixes the fragile turns, it does not level the models.

Which model should you choose?

Start with the voice: the model your callers will prefer may not be the most reliable, and this benchmark cannot choose it for you. Then make it reliable with the prompt, and change model only when the prompt is no longer enough.

On this script, if you prefer:

• GPT-Live-1: it asks almost every question from the start; the adjustment mostly makes it repeat less.
• Grok Voice 3.0: it makes the caller repeat the least; watch the end of the calls.
• Gemini 3.1 Flash Live: it gains the most from a well-written script; rewrite the four turns first.
• Gemini 3.8 Live: the adjustment moves it little; if the questions it skips matter to you, test another model.

Limits

• The voice is not measured: the benchmark counts questions; it does not judge how a model sounds.
• One simulated buyer, one language: a single simulated buyer, itself a model, and calls in French only.
• Few calls: 8 or 9 per model before the adjustment, 6 after; only two gaps are significant.
• Speech-to-speech only: the benchmark compares four speech-to-speech models with each other; no pipeline was measured on it.

Appendix: the adjusted prompts

Four snippets to paste as they are; replace the brackets with your questions. The CallShift skill carries them too.

1. A transition sentence before each block. In the rules of conduct:

Before each block of the script, a transition sentence, alone in its turn: what comes next and why. You ask the first question of the block on the next turn.

In the script, at the head of each block:

[Block name] (transition sentence first): "[Moving on to ..., because ...]"

2. Never a question deferred “to the next turn”. In the script, instead of “then, on the next turn: …”:

Q[n] bis. In its own turn, right after the answer to Q[n]: "[question]". It is a question of its own: you do not move on to Q[n+1] without asking it.

3. After a counter-question, answer, then ask again. In the rules of conduct:

If the person asks you a question, you answer in one sentence, then you ask your question again. Getting back on track means asking the current question again until it is answered, never moving on to the next one.

In the script, on a question with a choice:

If the person asks for details to choose, answer in one sentence, then: "So, which one do you pick?" You wait for the choice before moving on.

4. Always ask the question before announcing the task. In the script, on the question followed by a task:

Q[n]. "[question]" Always ask it, even if the person already mentioned it in passing.
TASK only after the answer to Q[n], never instead of Q[n]: half a sentence naming what you take care of.