Voice agents

• Best overallSGemini 3.8 Live1,088 on Artificial Analysis Speech Agent Arena (models)
• Best valueSGemini 3.8 LiveTier S for $0.01/min of audio
• Best free startSCartesia LineFree tier · from $5/mo · 1,034 on Artificial Analysis Speech Agent Arena (cascaded agent platforms)
• Best local (≤24 GB GPU)No qualifying option yet

• SABCTier: how close to the best result on the task's main public benchmark
• A Outlined: a tool inherits the tier of the best model it offers
• – No public result yet
• Cloud · Local: runs on your own hardware

CloudLocal

Models

The engines themselves. Use them through an API or inside the apps on the right.

Cloud

S Gemini 3.8 LiveGoogle · $0.01/min of audio1,088 on Artificial Analysis Speech Agent Arena (models)$S Grok Voice 3.0xAI · $0.08/min of audio1,070 on Artificial Analysis Speech Agent Arena (models)A GPT-Live 1OpenAI · $0.05/min of audio1,057 on Artificial Analysis Speech Agent Arena (models)B GPT-Realtime-2.1OpenAI · $0.19/min of audio941 on Artificial Analysis Speech Agent Arena (models)C Qwen3.5-Omni-PlusAlibaba (Qwen)861 on Artificial Analysis Speech Agent Arena (models)C GPT-Realtime-2.1 MiniOpenAI · $0.08/min of audio844 on Artificial Analysis Speech Agent Arena (models)C Qwen Audio 3.0 Realtime PlusAlibaba (Qwen) · $0.07/min of audio792 on Artificial Analysis Speech Agent Arena (models)– Higgs RealtimeBoson AINo public result yet– Nova 2.5 SonicAmazonNo public result yet– OpalDeepslate · $0.11/min of audioNo public result yet– StepAudio 3 RealtimeStepFunNo public result yet

Local

– LFM2.5-Audio-1.5BLiquid AI · ~4 GB VRAMNo public result yet– MiniCPM-o 4.5OpenBMB · ~21 GB VRAMNo public result yet– Moshi (moshiko)Kyutai · ~24 GB VRAMNo public result yet– NVIDIA-NemotronLabs-VoiceChat-11BNVIDIA · ~25 GB VRAMNo public result yet– PersonaPlex 7B v1NVIDIA · ~24 GB VRAMNo public result yet– Qwen3-Omni-30B-A3B-InstructAlibaba (Qwen) · ~80 GB VRAMNo public result yet– Raon-SpeechChat-9BKRAFTON · ~22 GB VRAMNo public result yet– Step-Audio R1.1StepFun · ~74 GB VRAMNo public result yet

Apps

Ready-made apps with a UI: sign in or install, then start.

Cloud

S GeminiGoogle · Free tier · $4.99/movia Gemini 3.8 Live$S GrokSpaceXSI (formerly xAI) · Free tier · $8/movia Grok Voice 3.0A ChatGPTOpenAI · Free tier · $20/movia GPT-Live 1– ClaudeAnthropic · Free tier · $20/moNo public result yet– Meta AIMeta · Free tierNo public result yet– Microsoft CopilotMicrosoft · Free tier · $9.99/moNo public result yet– SesameSesame AINo public result yet

Local

Nothing tracked yet

Voice-agent platforms

Build, test and deploy voice agents for phone lines and websites, with little code.

Cloud

S Cartesia LineCartesia · Free tier · $5/mo1,034 on Artificial Analysis Speech Agent Arena (cascaded agent platforms)$S ElevenLabsElevenLabs · Free tier · $6/mo998 on Artificial Analysis Speech Agent Arena (cascaded agent platforms)B Deepgram Voice Agent APIDeepgram · Free tier916 on Artificial Analysis Speech Agent Arena (cascaded agent platforms)B Retell AIRetell AI · Free tiervia GPT-Realtime-2.1– BlandBland AI · Free tierNo public result yet– Millis AIMillis AINo public result yet– PolyAIPolyAINo public result yet– SynthflowSynthflow AINo public result yet– Telnyx Voice AI AgentsTelnyxNo public result yet– ThoughtlyThoughtlyNo public result yet– UltravoxUltravox (Fixie) · Free tierNo public result yet– VapiVapi · Free tierNo public result yet– VoiceflowVoiceflowNo public result yet

Local

– BolnaBolna AI · Free tierNo public result yet

Frameworks

Open frameworks that wire speech recognition, a model and a voice into your own real-time agent.

Cloud

Nothing tracked yet

Local

– PipecatDaily (open source) · FreeNo public result yet– LiveKit AgentsLiveKit · FreeNo public result yet– speech-to-speech (Hugging Face)Hugging Face · FreeNo public result yet– TEN FrameworkTEN (Agora) · FreeNo public result yet– UnmuteKyutai · FreeNo public result yet

Everything, in order

All voice agents options in quality order

Tier	Name	Approach	Runs	Rated by	Price	Skill
S	Cartesia Line Cartesia	Voice-agent platforms	Cloud	1,034 on Artificial Analysis Speech Agent Arena (cascaded agent platforms)	Free tier · $5/mo	intermediate
S	Gemini 3.8 Live Google	Models	Cloud	1,088 on Artificial Analysis Speech Agent Arena (models)	$0.01/min of audio	—
S	Grok Voice 3.0 xAI	Models	Cloud	1,070 on Artificial Analysis Speech Agent Arena (models)	$0.08/min of audio	—
S	ElevenLabs ElevenLabs	Voice-agent platforms	Cloud	998 on Artificial Analysis Speech Agent Arena (cascaded agent platforms)	Free tier · $6/mo	beginner
S	Gemini Google	Apps	Cloud	via Gemini 3.8 Live	Free tier · $4.99/mo	beginner
S	Grok SpaceXSI (formerly xAI)	Apps	Cloud	via Grok Voice 3.0	Free tier · $8/mo	beginner
A	GPT-Live 1 OpenAI	Models	Cloud	1,057 on Artificial Analysis Speech Agent Arena (models)	$0.05/min of audio	—
A	ChatGPT OpenAI	Apps	Cloud	via GPT-Live 1	Free tier · $20/mo	beginner
B	Deepgram Voice Agent API Deepgram	Voice-agent platforms	Cloud	916 on Artificial Analysis Speech Agent Arena (cascaded agent platforms)	Free tier	intermediate
B	GPT-Realtime-2.1 OpenAI	Models	Cloud	941 on Artificial Analysis Speech Agent Arena (models)	$0.19/min of audio	—
B	Retell AI Retell AI	Voice-agent platforms	Cloud	via GPT-Realtime-2.1	Free tier	intermediate
C	Qwen3.5-Omni-Plus Alibaba (Qwen)	Models	Cloud	861 on Artificial Analysis Speech Agent Arena (models)	—	—
C	GPT-Realtime-2.1 Mini OpenAI	Models	Cloud	844 on Artificial Analysis Speech Agent Arena (models)	$0.08/min of audio	—
C	Qwen Audio 3.0 Realtime Plus Alibaba (Qwen)	Models	Cloud	792 on Artificial Analysis Speech Agent Arena (models)	$0.07/min of audio	—
–	Pipecat Daily (open source)	Frameworks	Local	No public result yet Cloud models: tier B	Free	advanced
–	LiveKit Agents LiveKit	Frameworks	Local	No public result yet Cloud models: tier A	Free	advanced
–	speech-to-speech (Hugging Face) Hugging Face	Frameworks	Local	No public result yet	Free	advanced
–	TEN Framework TEN (Agora)	Frameworks	Local	No public result yet	Free	advanced
–	Unmute Kyutai	Frameworks	Local	No public result yet	Free	advanced
–	Bolna Bolna AI	Voice-agent platforms	Local	No public result yet	Free tier	intermediate
–	Bland Bland AI	Voice-agent platforms	Cloud	No public result yet	Free tier	intermediate
–	Claude Anthropic	Apps	Cloud	No public result yet	Free tier · $20/mo	beginner
–	Higgs Realtime Boson AI	Models	Cloud	No public result yet	—	—
–	LFM2.5-Audio-1.5B Liquid AI	Models	Local	No public result yet	~4 GB VRAM	—
–	Meta AI Meta	Apps	Cloud	No public result yet	Free tier	beginner
–	Microsoft Copilot Microsoft	Apps	Cloud	No public result yet	Free tier · $9.99/mo	beginner
–	Millis AI Millis AI	Voice-agent platforms	Cloud	No public result yet	—	intermediate
–	MiniCPM-o 4.5 OpenBMB	Models	Local	No public result yet	~21 GB VRAM	—
–	Moshi (moshiko) Kyutai	Models	Local	No public result yet	~24 GB VRAM	—
–	Nova 2.5 Sonic Amazon	Models	Cloud	No public result yet	—	—
–	NVIDIA-NemotronLabs-VoiceChat-11B NVIDIA	Models	Local	No public result yet	~25 GB VRAM	—
–	Opal Deepslate	Models	Cloud	No public result yet	$0.11/min of audio	—
–	PersonaPlex 7B v1 NVIDIA	Models	Local	No public result yet	~24 GB VRAM	—
–	PolyAI PolyAI	Voice-agent platforms	Cloud	No public result yet	—	intermediate
–	Qwen3-Omni-30B-A3B-Instruct Alibaba (Qwen)	Models	Local	No public result yet	~80 GB VRAM	—
–	Raon-SpeechChat-9B KRAFTON	Models	Local	No public result yet	~22 GB VRAM	—
–	Sesame Sesame AI	Apps	Cloud	No public result yet	—	beginner
–	Step-Audio R1.1 StepFun	Models	Local	No public result yet	~74 GB VRAM	—
–	StepAudio 3 Realtime StepFun	Models	Cloud	No public result yet	—	—
–	Synthflow Synthflow AI	Voice-agent platforms	Cloud	No public result yet	—	beginner
–	Telnyx Voice AI Agents Telnyx	Voice-agent platforms	Cloud	No public result yet	—	intermediate
–	Thoughtly Thoughtly	Voice-agent platforms	Cloud	No public result yet	—	beginner
–	Ultravox Ultravox (Fixie)	Voice-agent platforms	Cloud	No public result yet	Free tier	intermediate
–	Vapi Vapi	Voice-agent platforms	Cloud	No public result yet	Free tier	intermediate
–	Voiceflow Voiceflow	Voice-agent platforms	Cloud	No public result yet	—	intermediate

Benchmarks used

• Artificial Analysis Speech Agent Arena (models) by Artificial Analysis (blind human votes, Elo from blind pairwise votes on live voice-agent conversations (overall slice; GPT-Realtime-1.5 anchored at 1000)) · snapshot 18 Oct 2026
  23 speech-to-speech model entries read from the speech-to-speech page's embedded arena results; the 4 vendor cascaded systems on the same board are in aa-speech-agent-arena-tools. No last-updated date shown.
• Artificial Analysis Speech to Speech Index by Artificial Analysis (automated tests, AA-Speech to Speech Index: weighted average of speech reasoning (Big Bench Audio), agentic performance (tau-Voice), arena preference and task success rate) · snapshot 18 Oct 2026
  19 models with all index components, read from the page's embedded arena results (field stsQualityIndexV2Score, which the index chart plots). Mixes automated tests with the arena's human votes. No last-updated date shown.
• Artificial Analysis Big Bench Audio (speech reasoning) by Artificial Analysis (automated tests, Accuracy on Big Bench Audio: spoken reasoning questions answered in audio (speech in, speech out)) · snapshot 18 Oct 2026
  41 model/provider entries with a score (the page states 41 models evaluated), including open full-duplex models (Moshi, PersonaPlex, Freeze-Omni). Measures reasoning, not voice quality or latency. No last-updated date shown.
• Artificial Analysis Speech Agent Arena (cascaded agent platforms) by Artificial Analysis (blind human votes, Elo from blind pairwise votes on live voice-agent conversations; entries are vendor default cascaded systems (STT + LLM + TTS platforms), ranks tools) · snapshot 18 Oct 2026
  Same board and Elo scale as aa-speech-agent-arena; only the platform entries: Cartesia Line (Default Cascaded System) 1033.6, ElevenLabs Agents (Default Cascaded System) 998.0, Deepgram Voice Agent (Default Cascaded System) 915.5, Inworld Realtime (Default Cascaded System) 876.2.
• Cekura Bench: voice agent platforms (workflow benchmark v1) by Cekura (automated tests, Reliability pass^3: share of 82 phone-call scenarios (appointment workflow with tools) a platform passes on all three runs, judged by Cekura's evaluator suite, ranks tools) · snapshot 12 Oct 2026
  Page 'Last updated Oct 12, 2026'; 'frozen v1 release', 8 configurations x 82 scenarios x 3 runs (246 retained calls each). Same table also ranks two speech-to-speech models (GPT Realtime 64.6, Gemini Live 30.5), recorded under cekura-voice-agents. Platforms chose their own LLM/STT/TTS configuration (published in github.com/cekura-ai/cekura-bench/provider-configurations), so results mix platform and model choices. Platform rows: Retell 75.6, LiveKit 70.7, ElevenLabs 69.5, Telnyx 68.3, Pipecat 63.4, Vapi 59.8; all tracked. Cekura sells voice-agent testing and observability, not voice agents; it integrates with the platforms it ranks.