Hosted conversational voice agents — state of the art, August 2026

Researched 06-08-2026 for the Guide page (studio). Prime criteria: interruption (barge-in) and fluidity (voice-to-voice latency, turn-taking), for a RAG-grounded, cloned-French-voice guide embedded on a static government web page. Facts labeled: [docs] = official documentation, [vendor] = official marketing/blog claim, [3rd-party] = external measurement. Companion report: open-stacks-06-08-2026.md.

1. ElevenLabs Agents (Conversational AI) — the incumbent

Interruption / barge-in

Detection is server-side, not client VAD: a proprietary turn-taking model + VAD decides barge-in; the client receives events and stops playback. The server streams a vad_score event (probability 0–1 the user is speaking) and, on interruption, an agent_response_correction event with the truncated transcript. [docs] Client events
Interruptions are enabled by selecting interruption as a client event in the agent's Advanced tab; disable for content that must be delivered in full. [docs] Conversation flow
New (6 July 2026): disable_interruptions replaced by interruption_mode with three states: allow · disable_during_tool · disable_during_tool_and_turn, with per-tool overrides. [docs] Changelog 6-07-2026
New (22 June 2026): interruption_ignore_terms (array of strings) — suppress turn detection on specific phrases (backchannels like "oui", "d'accord" won't cut the agent off). [docs] Changelog 22-06-2026
New (3 Aug 2026): vad config gains background_voice_detection — filters background speakers so they don't trigger false barge-ins. [docs] Changelog 3-08-2026

Turn-taking

turn_eagerness: eager | normal | patient. turn_timeout: 1–30 s. soft_timeout_config: filler audio ("Hmm… yes") when the LLM is slow — timeout 0.5–8.0 s, optional LLM-generated filler, up to 7 extra messages, randomize_fillers, max_soft_timeouts_per_generation 1–8. [docs] Conversation flow, Changelog 8-06, 22-06
New (8 June 2026): selectable turn model — turn_model: turn_v2 | turn_v3, default turn_v3, per agent or per workflow node. [docs] Changelog 8-06-2026
The turn system combines VAD with prosody/semantic signals ("read the meaning of what's being said to intuit when a turn has finished"); "speculative turn-taking" is in production. [vendor] Interaction models blog

Latency

Component figures: Flash v2.5 TTS ~75 ms model latency (explicitly excludes app + network; 32 languages incl. French) [docs] Models; Scribe v2 Realtime ASR ~150 ms (now the default ASR provider, replacing elevenlabs, since 8-06-2026) [docs] Models, Changelog 8-06; orchestration overhead "<100 ms" [vendor] Orchestration engine blog.
End-to-end: "sub-second pipeline" [vendor, no measured figure published] Interaction models. Independent benchmark of the TTS alone: ~255 ms median TTFB with network. [3rd-party] Deepgram analysis

Connection

JS SDK: voice conversations default to WebRTC, text-only to WebSocket; explicit connectionType: 'webrtc' | 'websocket'. WebRTC audio fixed at pcm_48000. Auth: public agents connect with just the agent ID (works from a static page); private agents need a server to mint a signed URL (WS) or conversation token (WebRTC). [docs] JS SDK, WebRTC token, WebRTC blog

New since June 2026 (beyond the above)

eleven_realtime_v1_mini is now the default realtime agent model (new realtime model family, 3-08-2026 — not yet on the public models overview page). [docs] Changelog 3-08-2026
Agent reasoning: enable_reasoning_summary, reasoning in transcripts, streamed via onAgentReasoningResponsePart (SDK). [docs] 7-06, 3-08
New knowledge_base system tool with enabled_strategies; RAG use_agent_defaults. [docs] 3-08
agent_response_complete WebSocket event (reliable end-of-turn detection client-side); pre_tool_speech (auto/force/off). [docs] Changelog April–June
Per-conversation cost_fiat (USD) + platform charge breakdowns in the API. [docs] 7-06

Pricing

All plans: overage $0.080/min; burst concurrency $0.160/min; LLM billed separately on top (deducted from credits). Included minutes: Free 15 · Starter 75 · Creator 275 · Pro 1,238 · Scale 3,738 · Business 12,375. [docs] Agents pricing
Third-party summaries put all-in cost at ~$0.08–0.12/min depending on LLM. [3rd-party] Cekura

2. OpenAI Realtime API

Current model: gpt-realtime-2.1 + gpt-realtime-2.1-mini (released ~6 July 2026), plus gpt-realtime-translate and gpt-live-transcribe. True speech-to-speech (no STT→LLM→TTS cascade). [docs] Realtime guide, community announcement
2.1 claims: better alphanumeric handling, better silence/noise handling, "more reliable interruption behavior when a user speaks over the model", p95 latency lowered ≥25% via caching, configurable reasoning effort. [vendor claims, no absolute ms figures published] DataNorth summary
Interruption: turn_detection = server VAD (threshold/silence) or semantic_vad — a classifier scoring the probability the user has finished speaking; less likely to cut users off mid-thought. [docs] VAD guide
Voices: built-in voices only — no custom or cloned voices; nothing in the 2.1 release changes this. [docs — absence of any cloning feature] Realtime docs. This remains the blocker for the guide's cloned voice.
Browser use: WebRTC for browsers, but a server must mint ephemeral credentials (POST /v1/realtime/client_secrets) — a static page alone cannot connect safely. [docs] Realtime guide
Pricing (official, token-based): 2.1 audio $32/1M in · $64/1M out (cached audio in $0.40); mini $10/1M in · $20/1M out. [docs] Pricing. Real-world: $0.18–0.46/min uncached; $0.05–0.10/min with caching (measured over 4,000 sessions); mini ≈ $0.016/min base math. [3rd-party] HackerNoon, eesel

3. Google Gemini Live API

Status: still preview. Models: gemini-2.5-flash-native-audio-preview-12-2025 and the newer gemini-3.1-flash-live-preview; "may change before becoming stable, more restrictive rate limits". [docs] Pricing page, Live overview
Voices: prebuilt Google voices only (Puck, Charon, Kore, Fenrir, Aoede, Leda, Orus, Zephyr) — no voice cloning; open feature requests confirm the gap. [docs + 3rd-party] Voice config, adk-docs issue #487
Barge-in: native — "users can interrupt the model at any time"; configurable automatic activity detection; affective dialog and proactive audio are its signature features. [docs] Live API, capabilities
70 languages incl. French; connection is a stateful WebSocket; ephemeral tokens recommended for client-side, but they are still minted server-side. [docs] Live API
Pricing: 2.5 native audio $3.00/1M audio-in · $12.00/1M audio-out; 3.1 flash live listed at $0.005/min audio-in · $0.018/min audio-out — by far the cheapest hosted option (~$0.02/min mixed). [docs] Pricing

4. New/notable hosted entrants (2026)

Cartesia Line: code-first voice-agent platform on Cartesia's SSM models; cloned voices yes, sub-90 ms TTS model latency [vendor], barge-in platform-managed; $0.06/min reported [3rd-party]. Line launch, Voximplant press
Deepgram Voice Agent API (GA): single API bundling STT (Nova-3/Flux) + LLM + TTS (Aura-2) with native barge-in and turn-taking prediction; bundled $4.50/hr = $0.075/min [vendor]; no self-serve arbitrary cloning; French TTS coverage weak (Aura-2 English-focused). Voice Agent API, GA post
Retell AI: telephony-first agent platform (web SDK exists); cloned voices via integrated ElevenLabs/other TTS; from ~$0.07/min [vendor]; measured default voice loop ~700 ms [3rd-party]. Pricing analysis, Softcery calculator
Vapi: orchestrator — $0.05/min platform fee + provider costs; cloned voices via ElevenLabs/Cartesia plug-ins; web SDK with public key (embeds without own server); measured default latency ~1,450 ms (config-dependent) [3rd-party]. Softcery, ainora comparison
Hume EVI: EVI 3 (EN/ES only) and EVI 4-mini (supports French); voice cloning from a sample and voice design from text [docs]; $0.04–0.07/min, 5 free min/mo [docs]. EVI versions, Voices, Pricing

5. xAI Grok Voice

Public API: yes. Grok Voice Agent API, OpenAI Realtime API-compatible (wss://api.x.ai/v1/realtime?model=… — OpenAI SDKs work by swapping the base URL). Models: grok-voice-think-fast-1.0, flagship grok-voice-think-fast-2.0; grok-voice-latest alias moves to 2.0 on 5 Aug 2026. [docs] Voice Agent API, LiteLLM provider page
Cloned voices: yes — "Clone any voice from a short reference clip with the Custom Voices API"; 80+ built-in voices. [docs] Voice Agent API
French: yes — 20+ languages "with native-quality accents", automatic language detection and mid-conversation switching. [docs] Voice overview
Interruption: server VAD only, threshold-based — turn_detection.threshold 0.1–0.9 (default 0.85), silence_duration_ms 0–10000, or manual. No semantic/prosodic turn model documented — a configuration generation behind ElevenLabs turn_v3 and OpenAI semantic_vad. [docs] Voice Agent API
Latency: "low-latency" [vendor, no published figures].
RAG: yes — file_search over uploaded document collections (vector stores), plus web search and X search server-side tools. [docs] Voice Agent API
Static page without server: no — ephemeral tokens recommended for browsers but minted server-side with the API key. [docs] Voice Agent API
Pricing: $0.05/min (1.0) · $0.08/min (2.0), voices included; tools billed per call (web/X search $5/1k, document search $2.50/1k); 100 concurrent sessions, 30-min max session. [3rd-party reporting of official rates] eesel, aipricing.guru. Also: no-code Voice Agent Builder beta since 1 July 2026 (phone-oriented). [3rd-party] eesel

Comparison table

Platform	Interruption quality	Voice-to-voice latency	Cloned voice	French	Web embed, no server	Price/min
ElevenLabs Agents	Best-configured: dedicated turn model (turn_v3), eagerness, `interruption_mode`, ignore-terms, background-voice filter	Components documented (75 ms TTS + 150 ms ASR + <100 ms orch.); "sub-second" e2e [vendor]	Yes (incl. professional clones)	Yes (32-lang Flash v2.5)	Yes (public agent widget/ID)	$0.08 overage + LLM (~$0.08–0.12 all-in)
OpenAI Realtime (gpt-realtime-2.1)	Good: server VAD + semantic VAD; "more reliable" in 2.1 [vendor]	Native s2s; p95 −25% in 2.1 [vendor]; no official ms	No (fixed voices)	Yes	No (server mints token)	Token-based; measured $0.05–0.46/min; mini ~$0.02
Gemini Live (3.1 flash live, preview)	Good native barge-in, auto VAD; still preview	Native audio, low; no official ms	No (8 prebuilt voices)	Yes (70 langs)	No (ephemeral token via server)	~$0.005 in + $0.018 out /min — cheapest
xAI Grok Voice (think-fast-2.0)	Basic: threshold server VAD only	"Low-latency" [vendor]; no figures	Yes (Custom Voices API)	Yes (20+ langs)	No (ephemeral token via server)	$0.05 (1.0) / $0.08 (2.0) + tools
Cartesia Line	Platform-managed barge-in	Sub-90 ms TTS model [vendor]; no e2e figure	Yes	Yes (Sonic multilingual)	Telephony-first; web unclear	~$0.06 [3rd-party]
Deepgram Voice Agent	Native barge-in + turn prediction [vendor]	Fast STT stack; no official e2e	No (stock Aura-2)	Weak (Aura-2 EN-focused)	No	$0.075 bundled
Retell	Provider-dependent	~700 ms measured [3rd-party]	Via ElevenLabs voices	Yes	No (server creates web call)	from ~$0.07 + costs
Vapi	Provider-dependent, configurable	~1,450 ms default measured [3rd-party]	Via ElevenLabs/Cartesia	Yes	Yes (public-key web SDK)	$0.05 + providers
Hume EVI 4-mini	Good (empathic turn model) [vendor]	No official figures	Yes (clone + design)	Yes (EVI 4-mini only)	No (token via server)	$0.04–0.07

Reading for the guide decision

ElevenLabs remains the only platform combining all four requirements: cloned voice + built-in RAG + French + serverless static-page embed. Its interruption stack is also the most configurable (turn_v3, interruption_mode, ignore-terms, background-voice filtering — all shipped June–Aug 2026).
The only credible challenger on the two prime qualities is OpenAI gpt-realtime-2.1 (native s2s fluidity, semantic VAD) — but it still has no cloned voices and needs a token server.
xAI Grok Voice is the notable 2026 newcomer: cloned voices + RAG + French + OpenAI-compatible API at $0.05–0.08/min, but its turn-taking is plain threshold VAD and it needs a server for browser tokens.
Gemini Live is the price disruptor (~10× cheaper) but is still in preview and has no cloned voices.

Hosted voice agents, the state of the art