Speech-to-speech APIs

8 hosted APIs on six dimensions. Every value links to its public source; a dash means the vendor does not publish it.

As of 2026-10-25


Grok Voice 3.0 xAI	66.2%	28	2 of 5	–	$0.08/min	16
gpt-realtime-2.1 OpenAI	49.4%	10	3 of 5	1.41 s	$64/1M out	–
Gemini 3.8 Live Google	42.1% Extended Thinking, high	30	2 of 5	–	$12/1M out	99
GPT-Live 1 OpenAI	–	31	3 of 5	0.80 s	$0.05/min	–
Eleven Realtime v1 ElevenLabs, cascaded	–	17,000+ library	1 of 5	–	$0.08/min	90+
Nova 2.5 Sonic Amazon	–	16	1 of 5	–	$12/1M out	7
Qwen-Audio 3.1 Realtime Plus Alibaba	–	13	2 of 5	1.92 s	$24/1M out	11
StepAudio 3 Realtime StepFun, preview	–	7	3 of 5	–	free preview	2

Instruction following

How faithfully the voice agent follows its instructions and policy across 213 customer-service scenarios (clean audio).

Share of customer-service calls completed under the policy prompt, across six languages.

No public benchmark scores all eight on the same setting yet. Rows carry the model variant each run used.

independent: EVA-Bench 2.0, τ-Voice v2

Response time

Time from the end of the user's turn to the model's reply. Each vendor ran the benchmark on its own model.

Not published: Gemini 3.8 Live, Grok Voice 3.0, Eleven Realtime v1, Nova 2.5 Sonic, StepAudio 3 Realtime

xAI and ElevenLabs publish response times measured by Artificial Analysis. Its terms do not allow republishing, so they are left out here.

vendor runs: OpenAI, Alibaba paper

Cost

List price per minute of session where the vendor charges by time. Extras apply; see notes.

Priced per 1M audio tokens, input and output

Gemini 3.8 Live	$3 / $12
Nova 2.5 Sonic	$3 / $12
Qwen-Audio 3.1 Realtime Plus	$6.4 / $24
gpt-realtime-2.1	$32 / $64

StepAudio 3 Realtime: Free during the preview; paid pricing not published.

GPT-Live 1. Per minute of session, billed per second. The backend reasoning model is billed separately at its own rates.
Grok Voice 3.0. Plus $0.004 per text input inside a session.
Eleven Realtime v1. Per minute of connected audio, plus LLM tokens: Mini $0.20 in and $6 out, Max $1.80 in and $9 out per 1M.
gpt-realtime-2.1. The mini variant is $10 in and $20 out per 1M audio tokens. Earlier turns are billed again as input on every turn.
Gemini 3.8 Live. Google also states $0.005 per minute of audio input and $0.018 per minute of audio output.
Nova 2.5 Sonic. US East (N. Virginia). Other regions cost more.
Qwen-Audio 3.1 Realtime Plus. International (Singapore) price list.

vendor pricing: GPT-Live 1, gpt-realtime-2.1, Google, xAI, ElevenLabs, Amazon, Alibaba, StepFun

Voices

Built-in voices selectable through the API, and whether you can add your own.

Eleven Realtime v1: 17,000+ library. Any voice in the ElevenLabs library, picked per session. Whether your own cloned voices work is not documented.

GPT-Live 1. Voice design from a text prompt is generally available. Cloning is limited to eligible customers.
gpt-realtime-2.1. Cloning is limited to eligible customers.
Gemini 3.8 Live. Uses the prebuilt voices of the Gemini TTS models. Custom voices are documented for the TTS models only.
Grok Voice 3.0. Cloning is available in the US only, excluding Illinois.
Nova 2.5 Sonic. AWS says eighteen; two voices are listed under both English (India) and Hindi.
StepAudio 3 Realtime. Cloned voices are made with StepAudio 2.5 TTS at $1.50 per voice.

vendor docs: GPT-Live 1, gpt-realtime-2.1, Google, xAI, ElevenLabs, Amazon, Alibaba, StepFun

Expressiveness

Documented controls over how the voice speaks. No public benchmark compares style following across these APIs yet.

	Style	Emotion	Speed	Accent	Nonverbal
GPT-Live 1		–			–
gpt-realtime-2.1		–			–
StepAudio 3 Realtime				–	–
Gemini 3.8 Live			–		–
Grok Voice 3.0					–
Qwen-Audio 3.1 Realtime Plus			–	–	–
Eleven Realtime v1	–			–	–
Nova 2.5 Sonic		–	–		–

documented not available– not documented

GPT-Live 1. Speed is set through the prompt; there is no speed parameter.
gpt-realtime-2.1. The speed parameter (0.25 to 1.5) changes playback rate only.
StepAudio 3 Realtime. All three go through a free-text instructions field that the docs say is not guaranteed.
Gemini 3.8 Live. The 3.8 Live model page lists affective dialog as removed; an older capabilities page still describes it.
Grok Voice 3.0. The prompting guide says instructions shape the words, not how the voice sounds. Speed runs 0.7 to 1.5.
Qwen-Audio 3.1 Realtime Plus. Emotion is an on/off switch (enable_speech_emotion); style goes through the instructions field.
Eleven Realtime v1. Speed control works inside ElevenAgents only. Expressive mode is managed by ElevenLabs and cannot be switched on.

vendor docs: GPT-Live 1, gpt-realtime-2.1, Google, xAI, ElevenLabs, Amazon, Alibaba, StepFun

Languages

Languages the model can hold a conversation in, as the vendor states.

Not published: GPT-Live 1, gpt-realtime-2.1

Gemini 3.8 Live. Other Google pages give 97 or 70.
Grok Voice 3.0. 20 locales covering 16 languages; xAI markets it as 20+.
Eleven Realtime v1. The ElevenAgents language setting lists 31; the launch post says 90+.
Qwen-Audio 3.1 Realtime Plus. Plus 21 Chinese varieties.

vendor docs: GPT-Live 1, gpt-realtime-2.1, Google, xAI, ElevenLabs, Amazon, Alibaba, StepFun

Artificial Analysis runs its own speech-to-speech leaderboard. Its terms do not allow republishing, so none of its numbers appear here, including those quoted in vendor posts. See their leaderboard.