On September 15, 2026, Google introduced two new voice AI models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. They are built to hold a real spoken conversation while doing work in the background — the same basic job as an AI receptionist that answers a phone, books a job, and follows up afterward. Google is not selling that as a finished business product. It released the underlying technology, and other companies are the ones who will build phone tools on top of it. Here is what actually changed, and what it means for the calls your business makes and takes.
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking through its Gemini API, Google AI Studio, and its own apps (Search, Workspace, the Gemini app). This is a developer and consumer release, not a packaged business phone system. The practical effect for a trade business: expect more AI voice agents calling and answering phones over the next year, built by Google and by other companies using models like this one, and expect customers to get more comfortable talking to one.
What Google actually released
Google's own announcement calls Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking its “most advanced live dialogue models yet.” They are native speech-to-speech models: the AI listens and answers in audio directly, instead of converting speech to text, thinking, and converting the answer back to speech in separate steps. Google's developer blog describes that older three-step approach as a “cascaded architecture” and positions the new models as a more streamlined alternative to it.
This is not Google's first attempt at real-time voice AI. It released an earlier version, Gemini 3.1 Flash Live, back in March 2026. Gemini 3.8 Live and its Extended Thinking variant are the direct successor, released about six months later. As of this writing on September 17, 2026, they are the newest voice AI models Google has confirmed.
What the new models can actually do
Based on Google's own description, three things stand out for anyone thinking about phone calls:
It keeps working while it keeps talking. Google says the model “executes tools and API calls in the background while continuing the conversation, so the model can acknowledge requests and keep chatting while tasks finish in the background.” In plain terms, it does not have to go silent to look something up or complete a step. The Extended Thinking version goes further, using short verbal cues like “Let me check that…” while it reasons through a multi-step task, then narrating its progress as it goes.
It follows a conversation across a lot of languages. Google says it “automatically detects and transitions between 97 supported languages mid-conversation,” and can ground its answers in whatever it is seeing through a camera or screen in near real time.
It is being measured against other voice AI, not just against itself. Google reports that Gemini 3.8 Live Extended Thinking ranked first on a third-party benchmark, Artificial Analysis’ Speech to Speech Quality Index, and that the standard Gemini 3.8 Live placed second on a separate leaderboard called the Speech Agent Arena. Those are industry benchmarks Google is citing, not an independent test we ran ourselves, and a lab score does not guarantee a good experience on a real, noisy phone call with a real customer.
A voice model that can talk and work at the same time is exactly the shape of AI that eventually answers — and makes — phone calls.
Who can actually use this today
This is the part that matters most for deciding how much attention to pay. As of today, Gemini 3.8 Live is aimed at three groups:
- Developers can build with it now through the Gemini API and Google AI Studio, at a published rate of $0.005 per minute of audio the model listens to and $0.018 per minute it speaks.
- Enterprises can access it in a private preview inside Gemini Enterprise, with a customer-experience version described as “coming soon.”
- Everyday users can talk to it inside Google's own products: Search Live, the Gemini app, and, for paying Workspace subscribers, inside Docs, Gmail, and Keep.
Notice what is missing from that list: a packaged product that answers a business's phone. Google did not announce one. What it announced is the engine other companies — AI receptionist tools, call centers, customer service platforms — can build with. A handful of voice platforms, including Agora, LiveKit, LangChain, Pipecat, and Vercel, already connect to it. The honest version of “Google launched a new voice AI” is: developers got a new tool today, and your customers may start noticing the results over the coming months, not necessarily this week.
What it means for the calls your business makes and takes
None of this means any specific AI receptionist now runs on Google's model — check with whatever tool you use, and do not assume it. What it does mean is broader.
More of the calls landing on your phone may be AI, not a person. Google already places automated calls to local businesses to check pricing and availability on a customer's behalf; we covered what happens when you don't answer one. Faster, more natural voice models make that kind of automated call cheaper for Google, and for other companies, to run at scale.
Your customers will get more used to talking to AI, and that cuts both ways. As people spend more time talking to a Live model in Search or Gmail, a well-run AI receptionist answering your business line will feel normal instead of jarring. The bar customers hold an AI voice to — natural pacing, actually understanding what was asked, not repeating itself — is being set by whichever voice AI Google and other large labs ship, whether or not your own phone system uses it.
What Loop does for your business does not change because of this announcement. Onvertz built Loop to answer your phone, book the appointment into your calendar, transfer urgent calls to a person, text a follow-up, and log every call as a transcript and summary. Which company's speech model powers any given tool is worth understanding, not a reason to chase every model announcement.
What to watch out for
A faster, more natural-sounding voice model is not automatically a more honest one. Keep three things in mind as more AI ends up on both ends of your business calls:
A model that sounds confident while quoting a price or promising an appointment slot is only useful if it is actually connected to your real calendar, pricing rules, and availability. Google's release describes what the model can do, not a guarantee that any specific business integration is correct.
AI audio is watermarked, but most callers will never check. Google says audio generated by its AI products, including these models, carries an inaudible SynthID watermark meant to help tools detect AI-generated speech and flag misinformation. Treat a suspiciously smooth voice on the phone the way you would treat a suspicious email: verify before you act on what it asked for.
A new-model announcement moves fast. Pricing, availability, and features on a model released days ago change quickly in its first months. Anything written about Gemini 3.8 Live today, including this article, is a snapshot; verify current details before making a purchasing decision based on it.
What to do about it now
You do not need to switch tools or chase this announcement. Two things are worth doing.
Test your own call path
Call your business number during a busy hour and again after closing. If a person or an AI receptionist answers, listen for whether it actually understands the question, or just sounds like it does.
Ask any AI phone vendor a direct question
What happens when a caller asks something outside its rules? A good AI receptionist should be able to say “I don't have that information, let me get you to someone who does” instead of guessing in a confident voice. That matters more than which company's model runs underneath it.
Google's release moves voice AI forward for developers and for its own products. It does not, by itself, change how your business should answer the phone — reachable, accurate, and honest about what it does and doesn't know still wins.