Voice AI agents in 2026: GPT-Live, Gemini Live and Voxtral 3

The voice models OpenAI and Google shipped in September 2026 let voice AI agents split talking from thinking. A full-duplex speech model holds the conversation and hands harder work to a reasoning model or to your tools. OpenAI’s GPT-Live 1 delegates to a backend you choose, and Google’s Gemini 3.8 Live Extended Thinking reasons in the background inside one model. Mistral’s Voxtral 3 voice agents, released on 20 October, run a managed cascade of transcription, a language model and speech.

If you already have a text agent, start with GPT-Live 1, at $0.05 a minute plus the backend’s tokens. Use a realtime voice API such as OpenAI’s Realtime API or Gemini 3.8 Live when one model should hear, reason and act, Voxtral 3 when audio must stay with an EU vendor, and a cascade when you need the text between steps.

We compared the options from each vendor’s documentation as of 26 October 2026, priced one 5-minute support call on each, and added lessons from a voice tutor we built on an AI study platform. We have not benchmarked these products, so latency figures are the vendors’ own claims.

Which voice agent stack to start with

As of 26 October 2026, this is where we would start.

For a product that already has a text agent, we would start with GPT-Live 1. The current version of our tutor runs on it, and the agent behind the voice stays testable in text. We would plan early for silence during hand-offs and duplicate answers, which that version ran into. For a buyer who needs an EU vendor, we would prototype on Voxtral 3 and commit once it leaves preview.

Speech-to-speech vs cascaded, and where delegation fits

OpenAI’s voice agents guide describes three voice agent architectures, and every product here fits one.² In a speech-to-speech design, one model hears the audio, decides what to do and replies in speech, as OpenAI’s Realtime API and Gemini 3.8 Live do. There are fewer hops, and the text is a by-product.

A cascade runs speech-to-text, a language model and text-to-speech in turn. OpenAI suggests it “when each stage needs to be visible or replaceable”, for example to store transcripts or run policy checks before the agent answers.² Voxtral 3, Deepgram’s Voice Agent API and the standard ElevenAgents pipeline are managed cascades. Mistral’s latency guide says each stage waits for the one before it, and that tool calls are often the longest wait.¹⁰

Delegation sits on top of either. GPT-Live 1 hands reasoning and tools to a backend in one of two modes, fixed per session. Responses delegation calls an OpenAI model you configure, and the guide suggests gpt-6-luna first and gpt-6-sol for harder tasks. Client delegation sends each request to your own agent.¹¹ Google keeps the split inside one model. Extended Thinking calls tools in the background while it speaks fillers, and every tool must be declared non-blocking.⁴ On plain Gemini 3.8 Live, a non-blocking function that forwards hard questions to a reasoning model can give a similar split.

A cascade still wins when you need the text for redaction, policy checks or audit logs, or a different model at each step. Deepgram’s agent can think with OpenAI, Anthropic, Google, Groq or AWS Bedrock models, or your own endpoint.⁸ It also suits compliance decided vendor by vendor.

GPT-Live vs Gemini Live vs Voxtral 3, as of 26 October 2026

The table summarises each vendor’s documentation as of 26 October 2026. “Not stated” means we found no figure. Prices are US list prices before tax, carrier charges and residency uplifts.

Google’s per-minute audio prices, $0.005 in and $0.018 out, match its token prices at about 25 tokens a second, but the Gemini Live API “charges you per turn for all tokens present in the session context window”.⁶³¹

What voice AI agents cost for a 5-minute support call

Per-minute and per-token prices are hard to compare by eye, so we priced one illustrative call on these assumptions.

• The call lasts 5 minutes, with 2 minutes of caller speech, 2 of agent speech and 1 of silence.
• It has 10 exchanges and 3 tool calls, so the model produces 13 responses.
• Instructions and tool definitions take 3,000 text tokens, plus 200 per tool call and result.
• OpenAI counts 600 audio tokens a minute for the caller and 1,200 for the agent,³² and Google about 25 a second.³¹
• Both APIs bill the conversation again on every response.³²³¹ On average each response re-reads 3,300 text tokens and half the call’s final audio, which is 1,800 audio tokens on OpenAI and 3,000 on Google.
• GPT-Live delegates 4 times, with 5,000 input and 800 output tokens each, uncached.

The totals leave out carrier charges, OpenAI’s 10% residency uplift, reasoning tokens, search fees and tax. As of 26 October 2026:

Most options land between about $0.19 and $0.40. The outlier is the Realtime API when re-reads miss the cache, which OpenAI applies automatically but calls best-effort. Changing instructions or tools mid-session lowers the cache rate.³²

Token billing also grows faster than call length, because each response re-reads a longer conversation. Stretched to 20 minutes on the same assumptions, the Realtime call costs about $1.00 fully cached and $13.47 with no cache hits. Gemini needs context compression past 15 minutes anyway, and Google presents it as a cost control too.¹⁸³¹

Voice agent latency and turn-taking

Only Mistral puts numbers on turn latency in the pages we read, claiming under one second per Voxtral 3 agent turn and under 500 ms for realtime transcription.⁷ OpenAI suggests measuring the wait for a useful spoken answer at the median and the 95th percentile.²

End-of-turn detection comes before anything else in a reply, so its wait counts in full.¹⁰ The Realtime API defaults to server_vad, which waits for silence, and offers semantic_vad, which judges whether the user has finished.¹⁵ Google recommends 500 to 800 ms of silence and warns that 100 to 200 ms splits one utterance into fragments.¹⁹

Every step before the first word adds to the wait. Our text assistant on the same platform chained three slow steps before each answer, and cutting them took its first word from 20 to 30 seconds to about 3.

Three lessons from our voice tutor

We built the tutor on OpenAI’s Realtime API and then GPT-Live 1, and our field note has the details and the numbers from our build notes.

End the turn when the audio drains. On the Realtime API, response.done marks the end of generation. In our tutor it arrived three to four times sooner than the audio finished playing, so end-of-turn logic fired while the tutor was still talking. Turns now end on output_audio_buffer.stopped, sent over WebRTC and SIP once the buffer drains. Over a WebSocket you track playback yourself, on GPT-Live too.¹² On Extended Thinking, turnComplete doesn’t mean the work is done, so Google says to wait for interaction_status to return to IDLE.⁴

Give response requests one owner. Only one response at a time can write to a Realtime session’s default conversation, and server-side VAD can start responses by itself. Six places in our code sent response.create with no guard. One guarded function now sends every request.

Limit pace and length separately. Our tutor’s model chained 5 visuals in 15 seconds, and the cap of 10 visuals per lesson that we added then ended lessons after about 3 minutes. A support agent sending texts or booking callbacks needs the same two limits, a minimum gap and a total.

What the agent says while a tool or the thinking model works

In the current version of our tutor, the voice goes silent until each hand-off to the teaching model is answered. Each vendor offers a way to fill that wait.

• GPT-Live 1. Speech and delegated work run independently, so a finished backend response doesn’t mean the caller heard the answer.¹¹ OpenAI suggests prompting the voice to acknowledge the request without guessing the result, and warns that a spoken “I’m checking that” doesn’t prove the work has started.³⁵
• Gemini 3.8 Live. A blocking function makes the model wait. A non-blocking one lets it keep talking, and the scheduling field decides whether the result interrupts, waits until idle or stays silent.³⁶ Extended Thinking speaks fillers such as “Checking flight options now”.⁴
• Voxtral 3. A short clip plays while a tool runs, which Mistral says “makes the wait easier to bear but doesn’t shorten it”. Client-side tools fail on phone lines, where none of your code runs.¹⁰²¹
• ElevenAgents. An optional filler plays after a wait you set between 0.5 and 8 seconds, with 3 seconds recommended.²⁹
• Deepgram. Function calls dispatch before the caller’s turn is confirmed unless defer_until_eot holds them, which Deepgram advises for side effects such as booking an appointment or charging a card.⁸

Beyond dead air, test for these failures on any stack.

• A repeated action. OpenAI asks you to track each action with an operation ID and check it before retrying, because “a lost response should not cause a second booking”.¹¹
• Orphaned work. Interrupting GPT-Live leaves backend work running until your application finishes or cancels it.¹¹ Gemini discards pending function calls on interruption and reports their IDs, so reconcile anything your code already started.¹⁹
• Duplicate answers. Our tutor once answered a single question four times, so track which questions have been answered and drop repeats before they are spoken.

Telephony, recording consent and AI disclosure

On telephony, GPT-Live can also place outbound SIP calls once outbound calling is enabled for your organisation.³⁷ ElevenAgents adds no telephony fee, and your carrier bills you directly.²⁸

The EU options carry conditions. OpenAI needs approval for abuse monitoring controls and a Modified Retention amendment, and adds a 10% uplift on models released from 5 March 2026.³⁸ Google Cloud’s Agent Platform, which now includes Vertex AI, lists Gemini 3.8 Live for EU multi-region processing. That endpoint excludes the UK and Switzerland.⁵³⁹ Deepgram’s EU endpoint keeps managed OpenAI and Google models in the EU, and other managed providers have no EU endpoint yet.⁹ ElevenLabs’ EU residency is for Enterprise customers, and processing can leave the region without Zero Retention Mode.³⁰ Google uses free-tier Live API data to improve its products, so use the paid tier for real calls.⁶

Recording consent depends on where the caller is. US federal law allows recording when one party consents, while California requires the consent of all parties to a confidential communication, with fines of up to $2,500 per violation.⁴⁰⁴¹ If you keep audio or transcripts, ask in the greeting.

AI disclosure applies to voice too. Article 50 of the EU AI Act has applied since 2 August 2026. People must be told they are interacting with an AI at the latest at the first interaction, unless it is obvious, and the Commission says that exception should be read restrictively.⁴²⁴³ In the US, the FCC ruled on 8 February 2024 that AI-generated voices in calls are “artificial” under the Telephone Consumer Protection Act, which brings them under its consent rules.⁴⁴ Many state chatbot laws starting in 2027 add disclosure duties, covered in US AI chatbot laws starting in 2027. GPT-Live may paraphrase a scripted greeting, so play the disclosure from your application.³⁵ This is an engineering reading, so check your own case with a lawyer.

A checklist for choosing and testing a voice agent stack

• Write down call length, talk time on each side, tool calls per call and where callers are.
• Price a typical call and your longest call on each candidate.
• Check each session limit against your longest call.
• Decide what the voice says during every wait, and how long silence may last.
• Give each action an operation ID, check it before retrying, and hold irreversible ones until the caller’s turn is confirmed.
• End turns when playback ends, and route every response request through one function.
• Test barge-in mid-sentence, mid-tool-call and mid-filler.
• Play the AI notice and any recording question from your application.
• Measure the median and 95th percentile time to first sound and to the useful answer on real phone audio, after every model or prompt change.

Frequently asked questions