Realtime voice AI models: seven you can actually talk to

Prices from official vendor price lists · October 2026
7 models with speech in and speech out at low enough latency for live conversation. Output prices run from $2.4 to $24 per 1M tokens across the 3 models with a paid per-token price; 2 have a free tier and 1 ships open weights.
Realtime voice
Services found: 7
#AI modelCapabilitiesAPI $/1M (in / out)Context windowPrice / moOutput $/1MOverviewVisit site
1

GPT-Realtime-2.1

OpenAI
Text & chatVisionAudio
$4 / $24128K tokensAPI only24View
2

GPT-Realtime-2.1 Mini

OpenAI
Text & chatVisionAudio
$0.6 / $2.4128K tokensAPI only2.4View
3

Gemini Omni Flash

Google
Text & chatVisionAudio
$1.5 / $9API only9View
4

Gemini 3.5 Transcribe Live

Google
AudioAdjustable thinkingSpeech-to-text
API onlyView
5

Nova 2 Sonic

Amazon
Text & chatAudioRealtime voice
API onlyView
6

Nova Sonic

Amazon
Text & chatAudioRealtime voice
API onlyView
7

Nemotron 3 Nano Omni

NVIDIA
Text & chatVisionAudio
$0 / $0256K tokensAPI only0View

Browse by need

Reasoning models169 modelsCoding assistants17 modelsVision and images156 modelsLong context 1M+82 modelsFree tier available80 modelsOpen weights88 models

Seven models support speech in and speech out at latency low enough for natural conversation - voice agents, phone support, live interpreting. GPT-Realtime-2.1 and its Mini variant lead, alongside Gemini Omni Flash.

Pricing here is unlike anything else in the rating: audio tokens bill separately from text and cost several times more, with GPT-Realtime-2.1 at $32 input and $64 output per million audio tokens. A minute of two-way conversation consumes a surprising number. Mini variants cut that roughly threefold and suit scripted flows. Where you need one direction only, speech-to-text is dramatically cheaper.

Audio tokens dominateAt $32 in and $64 out per 1M, voice economics look nothing like text.
Latency is the productSub-second turn-taking is what separates these from transcription plus TTS.
Mini variants existRoughly threefold cheaper and sufficient for scripted flows.

Models with this capability

The 7 models below all support this capability. Prices are input and output per 1M tokens from the vendor official price list; for dirhams multiply by 3.6725. Tap a card to open the model page.

GPT-Realtime-2.1
#1
OpenAI
in $4out $24128K tokens
GPT-Realtime-2.1 Mini
#2
OpenAI
in $0.6out $2.4128K tokens
Gemini Omni Flash
#3
Google DeepMind
in $1.5out $9
Gemini 3.5 Transcribe Live
#4
Google DeepMind
price on request
Nova 2 Sonic
#5
Amazon
price on request
Nova Sonic
#6
Amazon
price on request
Nemotron 3 Nano Omni
#7
NVIDIA
in $0out $0256K tokens

How to choose

  • Budget by conversation minute, not by request.
  • Check barge-in support — natural conversation requires interruption handling.
  • Keep the system prompt short; it is resent as context on a live session.
  • Fall back to speech-to-text where only one direction is needed.

Common mistakes

Long system prompts on voiceContext accumulates fast in a live call and bills at audio rates.
No call length capAn unattended session can run until the budget is gone.

Who this suits

Support linesReplacing IVR trees with actual conversation.
Field workHands-free interaction where typing is impractical.
AccessibilityVoice-first interfaces for users who need them.

Related filters

Speech-to-textTranscription with diarisation and timestamps
Audio understanding25 models that interpret speech and sound
Low latency73 small, fast models for volume
Priority tier65 models with a paid fast lane
All 239 modelsThe full rating with filters and prices

Frequently asked questions

How expensive is a voice minute?

Audio tokens run $32 input and $64 output per 1M on GPT-Realtime-2.1. A two-way minute consumes a surprising number.

Are the Mini variants good enough?

For scripted flows, usually yes — roughly threefold cheaper.

What if I only need one direction?

Speech-to-text is dramatically cheaper than a full duplex model.

Can it interrupt and be interrupted?

The better implementations support barge-in; test it, it is not universal.

Is latency good enough for phone calls?

On the leading models yes, provided your network path is short.

Frequently searched

By capabilityReasoningCodingVisionLong context 1M+Free tierOpen weightsLow latencyRealtime voiceImage generation
By vendorAnthropic (13)OpenAI (30)Google DeepMind (17)xAI (8)Meta (6)DeepSeek (9)Mistral AI (11)Alibaba (Qwen) (34)Z.ai (12)Moonshot AI (5)
Popular modelsClaude Fable 5.1GPT-6 AstraClaude Opus 5Claude Fable 5GPT-5.6 SolGLM 5.3Grok 4.6Kimi K3GPT-5.6 Terra
By deliveryBatch APIPrompt cachingPriority tierFrontier tierProprietary APITool callingStructured JSONWeb searchMultilingual

Ratings 0

Avg. score 0

Rate this page

Help us improve