Google shipped two new voice models today, and the big change has nothing to do with how they sound. By the end of this piece, you’ll know what you can do with them that you couldn’t do last week, what an hour of real conversation actually costs (the number going around is misleading), and whether to try it now or wait.
Verdict: Gemini 3.8 Live and 3.8 Live Extended Thinking are the first mainstream voice models that keep talking to you while they do work in the background. That’s a real change in how you use AI, and it’s cheap enough to try today. Two catches: the “$1.38 per hour” figure you may have seen assumes both of you talk nonstop for a full hour, and the Extended Thinking version charges extra for its behind-the-scenes reasoning. Try it now. Budget with the numbers below.
The hold-music problem
You know the drill with every voice assistant you’ve ever used. You ask for something. It goes quiet. Maybe a spinner appears, maybe nothing does, and you sit there wondering if it even heard you. Then it comes back with an answer, and you start over.
I call that the hold-music problem. You talk, it stops, it thinks, it does the task, then it talks again. Every request is a tiny customer-service call where the agent puts you on hold to go ask a manager. The hold music is silence.
What Google announced today removes the hold music. The model can say “Let me check that,” keep the conversation going, and finish the job behind the scenes. Google’s own demos have it narrating its progress out loud while a multi-step task runs.
That sounds small until you picture it.
“Check the flights, compare my calendar, message Sarah, and while you’re at it, tell me what I should know about the destination.”
The assistant keeps talking about the destination while the other three jobs run. You don’t wait. You don’t stare at a spinner. You keep having a conversation with something that is also working.
What you can do now that you couldn’t last week
Here is where this stops being a press release and starts being useful. All of these are things Google showed today, not things I’m guessing at.
Hand off a chain of tasks by voice and keep talking. Booking a trip used to mean one question, one wait, one answer, repeat. Now you can give the whole list at once and stay in conversation while it works through them. Google’s demo shows multi-step bookings and background lookups happening without a single pause in the chat.
Point your camera at something and ask about it, live. The model sees what’s on your screen or in front of your phone in near real time. Google’s examples include walking a new employee through an onboarding screen and narrating a chess game as the pieces move. For a tradesperson, that could mean pointing at a breaker panel and asking what’s wrong.
Switch languages mid-sentence without telling it to. It detects and moves among 97 languages on its own. A bilingual family or a front desk that serves tourists no longer needs a “switch to Spanish” button.
Watch it think out loud. With Extended Thinking, the model reasons and speaks at the same time. Instead of a silent pause before a hard answer, you hear it work: “Let me check that,” then a running commentary as it finishes. That changes how much you trust it, because you can hear where it’s going before it gets there.
Sketch on paper, describe it, get working software. One of Google’s demos turns rough sketches plus spoken feedback into working web components. You describe what you want while it builds, and you correct it by talking.
Try it without a developer. Extended Thinking is rolling out today inside the regular Gemini Live app for everyone. Google AI Pro and Ultra subscribers also get it in Google Docs, and all Google AI subscribers get it in Gmail and Keep. If you build things, both models are live in the Gemini API and Google AI Studio.
Every one of these was either impossible or clunky before today, because the model had to stop talking to do any of it.
What Google says it scored (and how to read that)
Google reports the Extended Thinking model in first place on Artificial Analysis’ Speech to Speech Quality Index with a score of 82.6, and leading two tests of whether a voice agent can actually finish a task: 68.6% on one called τ-Voice and 35.1% on a banking-task test from Sierra. Other coverage puts the cheaper base model at 76.0 on the same quality index.
Two things to keep in mind. Those charts were published by Google, even though they cite an independent tracker, so treat them as “Google says” until you can read them on the tracker’s own page. And OpenAI claims first place on a related test, called Tau3, at 86.2%. Different test, different scoring, so both companies can say “first” without lying. The plain read: the two are close on getting tasks done, and Google is cheaper.
Where OpenAI still has the edge
Don’t let this turn into “Google wins voice.” OpenAI’s GPT-Live-1 became available to developers on September 10, five days before this launch. Its strength is that it listens and talks at the same time, so you can interrupt it the way you’d interrupt a person. OpenAI reports a big jump on its own interruption test (80.1% versus 45.4% for its previous model), and early users say it still feels more natural in back-and-forth. That is a feel, not a measurement, so weigh it as one.
What it costs to run
Both Gemini 3.8 Live models are closed, Google-owned models. You can’t download them or run them yourself. They cost the same, per Google’s pricing page on September 15, 2026:
Your voice going in: $0.005 per minute
Its voice coming out: $0.018 per minute
Camera or screen video going in: $0.002 per minute
The model’s behind-the-scenes reasoning (Extended Thinking only): billed as text, $4.50 per million tokens
There’s a free tier in Google AI Studio, with the usual catch that Google uses your free-tier conversations to improve its products. Paid usage is not used that way. Big-company access through Gemini Enterprise is still an invitation-only preview.
The cost most people miss is that reasoning line. When the Extended Thinking model “thinks” while it talks, that thinking is billed separately from the audio. Google doesn’t say how much thinking a typical turn uses, so the only way to know your real bill is to run it on your own conversations and look.
Claims vs. Receipts
Claim: “Roughly $1.38 for an hour of conversation.” Receipt: that figure is 60 minutes of your voice at $0.005 plus 60 minutes of its voice at $0.018. It assumes you both talk for the entire hour, which never happens. A normal conversation where you each talk about half the time is closer to $0.69 per hour at Google’s published rates, before any reasoning charges. Both numbers come from Google’s pricing page, September 15, 2026.
Claim: “At least $3 per hour for OpenAI’s GPT-Live-1.” Receipt: OpenAI charges $0.05 per minute for the whole session, which is $3.00 an hour, and the clock keeps running during silence and hold time. On top of that, a second model does the hard thinking and bills its own usage. An independent write-up from TokenCost (September 11, 2026) puts a real working agent at roughly $3.05 to $5.16 per hour depending on which thinking model you pick. So “at least $3” is accurate, and a little kind to OpenAI.
Claim: Google’s chart shows GPT-Live-1 costing $5.83 per hour. Receipt: that’s Google’s chart, citing Artificial Analysis, relayed by OfficeChai on September 15. TokenCost’s own example lands at $5.16 for the same OpenAI setup. Same ballpark, different assumptions about how the conversation goes. Neither is wrong.
The bigger shift
Chat makes you wait. Voice that works in the background lets the job run underneath a conversation you’re already having. Whoever nails that for one specific line of work, with the right tools plugged in, is going to feel like magic to the person using it.
So What
Do this: Open Gemini Live on your phone today and hand it a three-step request to see the difference for yourself. If you build things, set up a 3.8 Live test in AI Studio with one real task behind it, and check the reasoning charges before you quote anyone a price.
Skip this: Tearing out an existing OpenAI voice setup because of today’s scores. The two are close, the tests don’t line up, and your users care more about how natural it feels than about a leaderboard.
Wait on this: Big-company deployment through Gemini Enterprise, which is still invitation-only, and Google Workspace access for business accounts, which Google lists as “coming soon.”
Steal this line: The next AI interface isn’t chat. It’s talking to an agent while it works.
Receipts/Sources:
The hold-music problem / What you can do now / What Google says it scored
Google, “Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking” (September 15, 2026): primary source for the launch, background task execution, 97 languages, live visual input, the “Let me check that” cues, the onboarding, chess, booking, and sketch-to-software demos, the 82.6 / 68.6% / 35.1% figures, and who gets access where.
OfficeChai, “Google Releases Gemini 3.8 Live-Extended Conversational Model, Claims Better Performance Than Rivals At Lower Price” (September 15, 2026): backs the 76.0 score for the base model and the competitor figures on Google’s charts.
Where OpenAI still has the edge
TokenCost, “GPT-Live-1 Pricing: $0.05/min Plus the Backend Bill” (September 11, 2026): backs GPT-Live-1’s September 10 release, its listen-while-talking design, the 80.1% vs 45.4% figures, and the 86.2% Tau3 claim.
What it costs to run / Claims vs. Receipts
Google AI for Developers, “Gemini Developer API pricing” (accessed September 15, 2026): primary source for every Gemini 3.8 Live price above and the free vs paid data terms.
TokenCost, “GPT-Live-1 Pricing: $0.05/min Plus the Backend Bill” (September 11, 2026): backs the $3.00 per hour session rate, silence billing, and the $3.05 to $5.16 per hour working range.