Analytics

Hosted speech-to-speech APIs side by side, as of 2026-10-25. Every value comes from a public source. An API without a public number is left out of that chart and listed under it.

GPT-Live 1gpt-realtime-2.1Gemini 3.8 LiveGrok Voice 3.0Eleven Realtime v1Nova 2.5 SonicQwen-Audio 3.1 Realtime PlusStepAudio 3 Realtime

Instruction following

EVA-Bench 2.0, faithfulness. How faithfully the voice agent follows its instructions and policy across 213 customer-service scenarios (clean audio). Whiskers show the reported range.

Higher is better

Full-duplexCascade

66.2%

49.4%

42.1%

Grok Voice 3.0

gpt-realtime-2.1

Gemini 3.8 LiveExtended Thinking, high

Source: EVA-Bench 2.0, 2026-10-15Not published: GPT-Live 1, Eleven Realtime v1, Nova 2.5 Sonic, Qwen-Audio 3.1 Realtime Plus, StepAudio 3 Realtime

Task success

τ-Voice v2, task success. Share of customer-service calls completed under the policy prompt, across six languages.

Higher is better

Full-duplexCascade

79.9%

68.5%

GPT-Live 1with GPT-6 Astra backend

Gemini 3.8 LiveExtended Thinking, high

Source: τ-Voice v2 leaderboard, 2026-10-20Not published: gpt-realtime-2.1, Grok Voice 3.0, Eleven Realtime v1, Nova 2.5 Sonic, Qwen-Audio 3.1 Realtime Plus, StepAudio 3 Realtime

Response time

Full-Duplex-Bench v1, turn-taking latency. Time from the end of the user's turn to the model's reply. Each vendor ran the benchmark on its own model.

Lower is better

Full-duplexCascade

0.80 s

1.41 s

1.92 s

GPT-Live 1

gpt-realtime-2.1

Qwen-Audio 3.1 Realtime Plus

Source: OpenAI, 2026-09-10; OpenAI, 2026-09-10; Alibaba, 2026-09-24Not published: Gemini 3.8 Live, Grok Voice 3.0, Eleven Realtime v1, Nova 2.5 Sonic, StepAudio 3 Realtime

Price per session minute

For APIs priced by connected minute.

Lower is better

Full-duplexCascade

$0.05

$0.08

$0.08

GPT-Live 1

Grok Voice 3.0

Eleven Realtime v1

Source: each vendor's pricing page, 2026-10-25Not published: gpt-realtime-2.1, Gemini 3.8 Live, Nova 2.5 Sonic, Qwen-Audio 3.1 Realtime Plus, StepAudio 3 Realtime

Price per 1M audio tokens

For APIs priced by audio token. Input and output are priced separately.

Lower is better

InputOutput

$3 / $12

$3 / $12

$6.4 / $24

$32 / $64

Gemini 3.8 Live

Nova 2.5 Sonic

Qwen-Audio 3.1 Realtime Plus

gpt-realtime-2.1

Source: each vendor's pricing page, 2026-10-25Not published: GPT-Live 1, Grok Voice 3.0, Eleven Realtime v1, StepAudio 3 Realtime

Languages

Languages the vendor lists for speech in and out.

Higher is better

Full-duplexCascade

99

90+

16

11

7

2

Gemini 3.8 Live

Eleven Realtime v1

Grok Voice 3.0

Qwen-Audio 3.1 Realtime Plus

Nova 2.5 Sonic

StepAudio 3 Realtime

Source: each vendor's documentation, 2026-10-25Not published: GPT-Live 1, gpt-realtime-2.1

Preset voices

Voices available without cloning. The label under each API shows custom-voice support.

Higher is better

Full-duplexCascade

31

30

28

16

13

10

7

GPT-Live 1Clone + design

Gemini 3.8 LiveCustom voices not stated

Grok Voice 3.0Clone

Nova 2.5 SonicNo custom voices

Qwen-Audio 3.1 Realtime PlusClone

gpt-realtime-2.1Clone

StepAudio 3 RealtimeClone

Source: each vendor's documentation, 2026-10-25. ElevenLabs offers a library of 17,000+ voices instead of presetsNot published: Eleven Realtime v1

Expressive controls

What the API documents for steering delivery. A filled dot is documented, an empty dot is documented as unavailable, and a dash means the vendor does not say.

API	Style	Emotion	Speed	Accent	Non-verbal
GPT-Live 1OpenAI		—			—
gpt-realtime-2.1OpenAI		—			—
Gemini 3.8 LiveGoogle			—		—
Grok Voice 3.0xAI					—
Eleven Realtime v1ElevenLabs	—			—	—
Nova 2.5 SonicAmazon		—	—		—
Qwen-Audio 3.1 Realtime PlusAlibaba			—	—	—
StepAudio 3 RealtimeStepFun				—	—

Source: each vendor's API documentation, 2026-10-25

Artificial Analysis results are not shown because its terms do not allow republishing them.