What is the best speech-to-speech AI model in 2026?

As of September 17, 2026, Gemini 3.8 Live Extended Thinking tops Artificial Analysis' Speech to Speech Index at 82.6, ahead of OpenAI's GPT-Live-1 paired with GPT-6 Astra (81.5) and GPT-Live-1 with GPT-5.6 Sol (80.1). Google's model leads agentic task completion (68.6% on τ-Voice, 35.1% on τ-Voice-banking) and speech reasoning (97.7% on Big Bench Audio). GPT-Live-1 remains the best pure full-duplex model for conversations with frequent interruptions. For cost-sensitive volume, plain Gemini 3.8 Live (second in the human-preference Speech Agent Arena) and GPT-Realtime-2.1-mini are the value picks.

How much do speech-to-speech models cost per minute in 2026?

Gemini 3.8 Live and 3.8 Live Extended Thinking: $0.005 per minute of audio in and $0.018 per minute out (Google's estimate from $3/$12 per million tokens), so roughly $0.01–0.02 per conversation minute. GPT-Live-1: $0.05 per minute for the voice layer plus a separately billed backend text model. GPT-Realtime-2.1: $32 per million audio input tokens and $64 per million output, about $0.05 per minute; GPT-Realtime-2.1-mini: $10/$20 per million, about $0.016 per minute. ElevenLabs Agents, a platform rather than a single model, is reported at roughly $0.08–0.10 per minute with the LLM bundled.

What is the difference between a speech-to-speech model and a voice pipeline?

A speech-to-speech (native audio) model takes audio in and produces audio out in one network, preserving tone, timing and interruptions — Gemini 3.8 Live, GPT-Live-1 and GPT-Realtime-2.1 all work this way. A voice pipeline, also called a cascade, chains speech-to-text, a text LLM and text-to-speech; it is more flexible (any LLM, any voice) but adds latency at each handoff and loses prosody. Platforms such as ElevenLabs Agents, LiveKit and Pipecat orchestrate either approach; Google's Live API partners (LiveKit, Pipecat, Agora, Vercel) now expose the native models directly.

Which speech-to-speech model should I use for a customer-support phone agent?

Gemini 3.8 Live Extended Thinking if the calls involve multi-step tasks (lookups, bookings, account changes) — it runs tool calls in the background while it keeps talking and has the best published τ-Voice scores. GPT-Live-1 if callers interrupt constantly and you want to choose your own reasoning model. GPT-Realtime-2.1 or its mini variant if you want one model, one bill and strict turn-taking. All three support telephony; none is open-weights, so there is no self-hosted option in this tier.

Quick Answer

Best Speech-to-Speech AI Models 2026: Top 6 Ranked

Published: September 17, 2026

The ranking at a glance

Rank	Model	Vendor	AA Speech to Speech Index	Price (audio)	Best for
1	Gemini 3.8 Live Extended Thinking	Google	82.6 (#1)	$0.005/min in · $0.018/min out	Complex agentic voice tasks
2	GPT-Live-1 (+ GPT-6 Astra or GPT-5.6 Sol backend)	OpenAI	81.5 (Astra medium) · 80.1 (Sol low)	$0.05/min + backend tokens	Full-duplex, interruption-heavy conversations
3	Gemini 3.8 Live	Google	2nd in Speech Agent Arena (index not published at launch)	$0.005/min in · $0.018/min out	Scale, cost, visual grounding
4	GPT-Realtime-2.1	OpenAI	Listed (turn-based)	$32/$64 per M audio tokens ≈ $0.05/min	One-model, one-bill voice agents
5	GPT-Realtime-2.1-mini	OpenAI	Listed (turn-based)	$10/$20 per M audio tokens ≈ $0.016/min	High-volume, budget IVR
6	ElevenLabs Agents (platform)	ElevenLabs	Not a native S2S model	≈ $0.08–0.10/min reported, LLM bundled	Voice quality, cloning, all-in-one build

Scores are from Artificial Analysis’ Speech to Speech Index, a weighted average of speech reasoning (Big Bench Audio), agentic performance (τ-Voice), human arena preference and task success rate. Verified September 17, 2026; this list will change — re-check the index before committing.

1. Gemini 3.8 Live Extended Thinking — best overall

Released September 15, 2026, Google’s high-complexity live model took the #1 spot on the Speech to Speech Index at 82.6, with 68.6% on τ-Voice, 35.1% on Sierra’s τ-Voice-banking and 97.7% on Big Bench Audio. Its signature trick is reasoning and speaking at once: it acknowledges with “Let me check that…”, narrates progress on multi-step background tasks, and finishes without dead air. It auto-switches among 97 languages, parses alphanumerics (claim numbers, confirmation codes) reliably, and accepts near-real-time visual input. Configurable thinking is exposed in the Live API.

Price: $0.005/min in, $0.018/min out — the same as plain 3.8 Live, which makes it the cheapest frontier voice model on this list per minute. Availability: Gemini Live API, Google AI Studio, private preview in Gemini Enterprise, and consumer surfaces (Gemini Live, Docs for AI Pro/Ultra, Gmail and Keep). Not open-weights.

2. GPT-Live-1 — best full-duplex conversation

OpenAI’s GPT-Live-1 shipped in ChatGPT in July 2026 and in the API on September 10, 2026. It is a true full-duplex model — it listens while it speaks — and OpenAI reports a 30-point lead over GPT-Realtime-2.1 on Full Duplex Bench. It delegates reasoning and tool calls to a backend text model of your choice (GPT-6 Astra, GPT-5.6 Terra or Luna, or third-party), which is why the AA index lists it per configuration: 81.5 with Astra (medium), 80.1 with Sol (low). Before Gemini 3.8 Live Extended Thinking, the Astra configuration was #1.

Price: $0.05/min for the voice layer, billed per second, plus backend tokens — trivial with Luna ($0.20/$1.20 per MTok), material with Astra ($10/$50). Supports telephony, ASR transcripts, keyword biasing and explicit turn detection.

3. Gemini 3.8 Live — best value at scale

The “scale and cost efficiency” sibling, also September 15, 2026. It placed second in the Speech Agent Arena, the blind human-preference test, and shares Extended Thinking’s 97-language switching, background tool calls and visual grounding — without the deep multi-step reasoning. At $0.005/$0.018 per minute it is the model to default to for high-volume assistants where a human rated it nearly as well as anything and the tasks are one or two steps deep.

4. GPT-Realtime-2.1 — best single-model, turn-based agent

Released July 6, 2026, GPT-Realtime-2.1 is OpenAI’s speech-to-speech model that does voice and reasoning in one network, turn-based. Billing is per audio token — $32 in / $64 out per million, cached input $0.40 — which OpenAI estimates at about $0.05 per conversation minute. Pick it when you want strict turn boundaries, one line item and no backend-model decision. As of September 2026 OpenAI positions GPT-Live-1 as the natural-conversation upgrade, not a replacement.

5. GPT-Realtime-2.1-mini — best budget option

The mini variant (July 7, 2026) costs $10/$20 per million audio tokens, about $0.016 per minute — a third of full Realtime-2.1 and comparable to Gemini 3.8 Live’s all-in rate. Quality is a step down, but for menu-style IVR, appointment confirmations and FAQ bots it is the cheapest OpenAI voice.

6. ElevenLabs Agents — best voices, not a native model

ElevenLabs Agents is an orchestration platform (STT, pluggable LLM, ElevenLabs TTS) rather than a single speech-to-speech network, so it does not appear on the native index. It earns a slot because for many products the voice is the brand: the largest voice library, best-in-class cloning, and an all-in-one builder. Reported pricing is around $0.08–0.10 per minute with the LLM bundled. Turn-taking is fast but not simultaneous.

How to choose

Multi-step tasks on the call (bookings, account changes, banking)? → Gemini 3.8 Live Extended Thinking.
Callers talk over the agent; you want your own reasoning model? → GPT-Live-1.
Millions of short calls, tight budget, human-rated quality still matters? → Gemini 3.8 Live or GPT-Realtime-2.1-mini.
One model, one bill, strict turns? → GPT-Realtime-2.1.
A specific brand voice or cloned voice? → ElevenLabs Agents (or ElevenLabs TTS behind a native model via LiveKit/Pipecat).
Need to self-host? → None of the above. The open-weights speech-to-speech tier trails the hosted frontier substantially in 2026; treat it as a separate category.

Three things the benchmarks do not tell you

Cost is dominated by talk time, not list price. A verbose agent on a cheap model can out-spend a terse one on an expensive model; measure minutes of output audio in your own pilot.
Full-duplex changes UX design. GPT-Live-1’s simultaneous listening means your prompts must handle backchannels (“mm-hm”) that turn-based models never see.
Watermarking and compliance differ. All Gemini audio carries a SynthID watermark; check each vendor’s data-retention and telephony-recording terms before deploying in regulated industries.