# New AI Voice Models Just Dropped. Should Your AI Receptionist Switch? (How We Test a Swap in 2026)

Author: Kris Hubbard
Published: October 12, 2026
Updated: October 12, 2026
Canonical: https://wrkless.com/blog/ai-receptionist-voice-models-2026

**Short answer:** Since September 10, 2026, OpenAI, ElevenLabs, Google, Alibaba, Microsoft, Decagon and Amazon have all shipped new voice models built for live conversation. The headline upgrade is full-duplex audio: agents that can listen while they talk, so callers can interrupt without breaking the call. That's real progress, but don't swap your AI receptionist's model in launch week. Test the new model on a staging copy, judge it on readbacks, bookings and transfers rather than how pretty it sounds, then move a slice of live calls over. The voice is the cheapest layer to change and the easiest one to break.

We build AI voice receptionists on Retell for service businesses, and every one of these launches lands in our inbox as the same question from owners: "Should we switch?" Here's how we answer it.

## What voice models launched in the last five weeks?

Here's the wave, oldest first, with the part that matters on a business phone line.

| Date (2026) | Release | What it is | Why a receptionist cares |
|---|---|---|---|
| Sep 10 | [OpenAI GPT-Live-1 in the API](https://openai.com/index/introducing-gpt-live-1-in-the-api/) | Full-duplex voice model that hands reasoning and tool calls to a backend model | Better interruptions; OpenAI lists telephony support |
| Sep 23 | [Gemini 3.8 Flash TTS and Flash-Lite TTS](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/) | Text-to-speech, with Flash-Lite aimed at high volume | More expressive voice in a standard stack |
| Sep 28 | [Eleven v4 and Eleven v4 Turbo](https://elevenlabs.io/blog/eleven-v4) | Text-to-speech; Turbo built for agents | ~150ms median time to first speech, 90+ languages |
| Sep 28 | [Qwen-Audio-3.1-Realtime](https://www.marktechpost.com/2026/09/28/alibaba-qwen-releases-qwen-audio-3-1-realtime-a-full-duplex-voice-model-trained-to-think-act-and-decide-when-to-speak/) | Full-duplex speech model with tool calling; no open weights announced | Another speech-to-speech option, API only |
| Oct 1 | [Microsoft MAI-Transcribe-2-Streaming and MAI-Voice-2.1](https://microsoft.ai/news/our-first-streaming-transcription-model/) | Streaming transcription in 60 languages plus new voices | Faster, cheaper ears and mouth for cascaded agents |
| Oct 1 | [Decagon Voice 3 and Chord](https://decagon.ai/blog/voice-3) | Duplex agent; voice model slows down for codes and phone numbers | Aimed at enterprise customer service |
| Oct 5 | [Amazon Nova 2.5 Sonic](https://aws.amazon.com/about-aws/whats-new/2026/10/amazon-nova-2.5-sonic/) | Speech-to-speech model, generally available in Bedrock | Better tool-calling accuracy, lower latency |

## What's the difference between full-duplex and the stack most AI receptionists run today?

Most AI receptionists today run a cascaded stack: speech-to-text hears the caller, a language model decides what to say, and text-to-speech says it. Each handoff adds delay, and the agent usually waits for a clean pause before it answers. OpenAI's own launch post makes this argument, saying GPT-Live-1 avoids "the latency and brittle handoffs of chained STT–LLM–TTS architectures."

Full-duplex models listen and speak at the same time. When a caller cuts in with "no, the other address," the agent can stop and adjust instead of finishing its sentence. OpenAI reports GPT-Live-1 beats its own GPT-Realtime-2.1 by 30 percentage points on Full Duplex Bench.

Two things the launch posts make clear, though:

- **Benchmarks are mixed.** Qwen's realtime model cut replies to background chatter, but MarkTechPost reports its interruption stop latency at 1.116 seconds versus 0.383 for GPT-Realtime-2. Every model wins on its own chart.
- **The voice layer doesn't run your business.** OpenAI's delegation guide says your application still "owns permissions, confirmations, business records, and task state." That's the half we spend most of our build time on: the webhook that writes the lead to the CRM, the calendar check, the transfer rules.

## Should you switch your AI receptionist to a new voice model right now?

Test now, switch later. Here's our reasoning.

On a receptionist line, the failures that cost money aren't about tone. They're the callback number read back wrong, the street address that lands in the CRM with a typo, the appointment that books twice, and the transfer that drops. A more emotive voice that mangles a unit number loses you the job.

You may also have switched models already without knowing it. Retell's [deprecation log](https://docs.retellai.com/deprecation-notice/overview) shows ElevenLabs Turbo voices auto-migrated to Flash on July 12, 2026, and Claude and Gemini models auto-migrated on May 25, 2026. Platforms do this to keep agents running, and it's mostly fine. But if nobody re-ran your test calls after those dates, nobody checked.

The newest models also aren't all on your platform yet. Retell's [pricing page](https://www.retellai.com/pricing) lists ElevenLabs v4 and v4 Turbo and the GPT Realtime 2.1 models, but not GPT-Live-1. Using GPT-Live-1 directly today means building your own telephony and orchestration, which is a much bigger project than flipping a dropdown.

## What does a better AI voice cost per minute?

Retell's list prices, checked October 12, 2026:

- Retell Voice Infra: $0.055/minute
- Retell Platform Voices (also Cartesia, OpenAI, MiniMax, Fish and Inworld voices): $0.015/minute
- ElevenLabs v4 Turbo: $0.08/minute
- ElevenLabs v4: $0.10/minute
- GPT Realtime 2.1 speech-to-speech: $0.38/minute
- GPT Realtime 2.1 mini: $0.07/minute

Going direct, OpenAI prices GPT-Live-1 at $0.05 per minute for the voice layer, billed per second, with the backend model charged separately.

Moving a Retell agent from a $0.015/minute voice to ElevenLabs v4 Turbo adds $0.065 per minute. At 3,000 call minutes a month, that's $195 a month. Measure that against what one extra booked job is worth to you, not against the old bill. If the new voice doesn't move bookings, it's a cost. If it keeps callers on the line, it pays for itself.

## How do we test a model swap before it touches a live phone line?

This is the process we run on our own Retell builds. It treats a voice change like a code change.

1. **Pin versions.** We never point a production agent at "whatever's newest." Retell's Agent Versioning 2.0 gives us staging and production tags, so the live number stays on a known version.
2. **Change one layer at a time.** Voice, or language model, or a full speech-to-speech swap. Never all three. Otherwise you can't tell what broke.
3. **Run the same scripted calls.** We reuse the test script from our [AI receptionist buying guide](/blog/ai-receptionist-small-business-buying-guide): a name that has to be spelled, a callback number, an address with a unit number, a caller who changes the appointment time mid-sentence, an after-hours emergency, and a caller who asks "is this a robot?"
4. **Check the side effects, not just the audio.** Did one booking land, not two? Did the webhook payload hit the CRM with the right fields? Did the transfer confirm? Retell's August changelog fixed agents firing the same tool call twice, like a booking confirmation. Re-run that check after any model change.
5. **Re-check your disclosures.** A new or cloned voice doesn't change your duty to [disclose that callers are talking to AI](/blog/ai-receptionist-disclosure-laws) or [handle recording consent](/blog/ai-call-recording-laws). ElevenLabs says every Eleven v4 voice clone requires verified consent from the voice's owner.
6. **Split live traffic.** Retell's A/B testing splits calls by percentage. We start the new version at about 20% of inbound calls.
7. **Watch a full week, then decide.** Compare booking rate, transfer rate, hang-ups and latency against the old version, and keep the one-click rollback ready.

If a model can't pass steps 3 and 4 on staging, we don't care how it sounds.

## Where does this wave actually change what we build?

**Interruption-heavy calls get better first.** After-hours seller calls and home-service emergencies are full of people talking over the agent. That's where full-duplex helps most, and it's where we point new models first in our [real estate after-hours setup](/blog/ai-receptionist-for-real-estate-after-hours-seller-calls) and [after-hours call answering](/after-hours-ai-call-answering.php) builds.

**The split architecture matches how we already build.** OpenAI's delegation model, with a fast voice layer up front and a smarter backend doing the work, is the same split we use: Retell handles the conversation, and the [webhooks and integrations](/retell-ai-integrations-and-webhooks.php) handle bookings and CRM writes. We covered where that logic should live in [Retell Workflows vs. n8n and Zapier](/blog/retell-workflows-vs-n8n-zapier). A better voice model doesn't replace that layer. It needs it more.

**Callbacks get more reliable.** Retell's August update fixed iOS and Google call screening being misread as voicemail, which had agents hanging up or leaving messages on live pickups. If your agent calls leads back, that fix matters more than any new voice.

**Multilingual gets cheaper to offer.** Eleven v4 covers 90+ languages and MAI-Transcribe-2-Streaming handles 60 with automatic language detection. For Orlando businesses with Spanish-speaking callers, that's a real upgrade path.

The model race will keep going. What a business owner should pay for is the setup around it: the test script, the version control, the CRM sync and the rollback plan. That's what keeps the next launch from breaking your phone line.

Want a second opinion on whether your receptionist should switch? See how we handle [Retell voice agent setup](/retell-ai-voice-agent-setup.php), or [book an AI systems assessment](https://cal.com/kris-hubbard-wrk-less/ai-systems-assessment) and we'll run your current agent through our swap test on the call.

## Questions people ask

### What is a full-duplex voice model?

A full-duplex voice model can hear and talk at the same moment, the way people do on a normal phone call. Older voice agents wait for the caller to stop, transcribe what they said, think, and then speak. Full-duplex models such as OpenAI's GPT-Live-1 and Alibaba's Qwen-Audio-3.1-Realtime can stop mid-sentence when a caller cuts in and change course.

### Will a newer AI voice make my receptionist book more appointments?

Only if it keeps callers on the line and captures their details correctly. A more natural voice can reduce hang-ups, but bookings depend just as much on accurate readback of names, numbers and addresses, single (not duplicate) calendar bookings, and working transfers. Compare booking rate and transfer rate on a slice of live calls before deciding.

### Can I use OpenAI's GPT-Live-1 for my business phone line today?

Yes, but not as a simple setting on most receptionist platforms yet. OpenAI offers GPT-Live-1 through its own Live API with telephony support, at a per-minute voice price plus backend model costs. As of October 12, 2026, Retell's public pricing lists GPT Realtime 2.1 and ElevenLabs v4 options but not GPT-Live-1, so using it directly means building your own call handling and integrations.

### Do AI receptionist platforms change the underlying model without telling you?

They can. Retell's deprecation log shows ElevenLabs Turbo voices were automatically moved to Flash voices on July 12, 2026, and older Claude and Gemini models were auto-migrated on May 25, 2026. Subscribe to your platform's deprecation notices and re-run your test calls after every forced migration.

### How long should I test a new voice model before switching every call?

Plan on about a week of live traffic after staging tests pass. Route roughly 20% of inbound calls to the new version, keep the old version on the rest, and compare bookings, transfers, hang-ups and latency across a full week so weekday and weekend callers are both represented. Keep a one-click rollback ready.

## Sources

- https://openai.com/index/introducing-gpt-live-1-in-the-api/
- https://developers.openai.com/api/docs/models/gpt-live-1
- https://developers.openai.com/api/docs/guides/live-delegation
- https://elevenlabs.io/blog/eleven-v4
- https://elevenlabs.io/v4
- https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/
- https://microsoft.ai/news/our-first-streaming-transcription-model/
- https://decagon.ai/blog/voice-3
- https://aws.amazon.com/about-aws/whats-new/2026/10/amazon-nova-2.5-sonic/
- https://www.marktechpost.com/2026/09/28/alibaba-qwen-releases-qwen-audio-3-1-realtime-a-full-duplex-voice-model-trained-to-think-act-and-decide-when-to-speak/
- https://www.retellai.com/pricing
- https://www.retellai.com/changelog
- https://docs.retellai.com/deprecation-notice/overview
- Retell pricing page, checked Oct 12, 2026: Retell Voice Infra $0.055/minute; Retell Platform Voices, Cartesia, OpenAI, MiniMax, Fish and Inworld voices $0.015/minute; ElevenLabs v4 Turbo $0.08/minute; ElevenLabs v4 $0.10/minute; GPT Realtime 2.1 $0.38/minute; GPT Realtime 2.1 mini $0.07/minute
- OpenAI GPT-Live-1 launch post, Sept 10, 2026: available in the API at $0.05 per minute for the front-end voice layer, billed per second; backend model and tool usage billed separately
- Retell changelog, March 19, 2026 (A/B Test Your Voice Agents): split call traffic by percentage across multiple agents, e.g. test a new prompt on 20% of inbound support calls
- Wrk Less calculation, Oct 12, 2026: $0.08/minute minus $0.015/minute adds $0.065 per minute; 3,000 minutes x $0.065 = $195 a month