Google just did something interesting with voice AI.
Instead of releasing one new model, they dropped two. And the split tells you a lot about where this technology is heading.
The models are Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. They sound similar. They're not. One is built for speed and scale. The other is built for thinking through hard problems while holding a conversation.
Here's what that actually means in practice.
Two Models, Two Very Different Jobs
Let's start with the split, because it matters.
Gemini 3.8 Live is the everyday model. It's designed for voice agents that handle lots of conversations without burning through your budget. Fast, efficient, good enough for most tasks.
Gemini 3.8 Live Extended Thinking is the heavy hitter. It reasons through multi-step problems while continuing to talk. It doesn't go silent while it thinks. It narrates its progress so you know it's still working.
The quality gap between them is about 6.6 points on benchmarks. That gap is the entire decision.
Think of it this way. One is a conversational interface. The other is a reasoning system that happens to speak.
What Gemini 3.8 Live Actually Does
Here's the practical stuff.
It processes audio directly. No separate transcription step. Speech goes in, speech comes out. That's faster and loses less nuance than the old approach of chaining together speech recognition, text processing, and speech synthesis.
It sees what you see. The model handles near real-time visual input through your camera or screen. Google's demo showed it narrating a chess game, watching the board and commenting as moves happened.
It switches languages automatically. 97 languages supported, and it can transition between them mid-conversation without you having to tell it.
It works in the background. The model acknowledges your request, keeps chatting, and runs API calls or tools behind the scenes. No awkward silence while it works.
That last point is more important than it sounds. Silence is the enemy of voice agents. Users don't want to wonder if the system crashed.
What Extended Thinking Adds
This is where things get genuinely new.
Extended Thinking reasons and speaks at the same time.
When you give it a complex task - debugging code, planning a trip, walking through a tax scenario - it doesn't go quiet while it figures things out. It uses verbal cues like "Let me check that…" to acknowledge your request, then narrates its progress as it works through the problem in the background.
It's using something Google calls an asynchronous reasoning protocol. The conversation keeps flowing while the model processes multi-step tasks that would normally require a pause.
That's a meaningful change. It means you can have a real back-and-forth with an AI that's actually thinking, not just responding.
How It Performs
Google released benchmark scores, and Extended Thinking came out on top in several categories.
On Artificial Analysis's Speech to Speech Quality Index, Extended Thinking took the number one spot. It beat GPT-Live-1 Astra and Grok Voice Think Fast 2.0.
On agentic task completion, it led the pack. On customer-service benchmarks, it led again. On reasoning tests, it scored 97.7%.
The standard Gemini 3.8 Live didn't top the charts. It ranked fifth on the Speech to Speech Index. But it ranked second on the Speech Agent Arena and first on ServiceNow's EVA-Bench.
Different tools for different jobs. That's the point.
What It Costs
Google is competing hard on price.
Gemini 3.8 Live costs $0.005 per minute of audio input and $0.018 per minute of audio output. That works out to roughly $1.38 to $1.50 per hour of conversation.
Extended Thinking costs more. Around $3.50 per hour for input audio. But that's still cheaper than Grok Voice Think Fast 2.0 at $4.80 and GPT-Live-1 Astra at $5.83.
The message is clear. Google is willing to undercut its rivals to get developers building on its voice stack.
How to Access It
For developers: Both models are in the Gemini API and Google AI Studio. The model codes are gemini-3.8-live and gemini-3.8-live-extended-thinking.
For enterprises: Both are in private preview in Gemini Enterprise, with plans to expand to customer experience use cases.
For regular users: Gemini 3.8 Live is available in Search Live. Extended Thinking is in Gemini Live. Google AI Pro and Ultra subscribers can also use Extended Thinking in Google Docs, Gmail, and Keep.
Google is also partnering with platforms like LiveKit, Pipecat, Agora, and Vercel to make integration easier.
Why This Matters
Voice AI has been stuck for years.
The old systems were either fast but dumb, or smart but slow. You couldn't have both. Ask something simple and you'd get a snappy response. Ask something complex and the system would either give a shallow answer or go silent while it "thought."
Gemini 3.8 Live breaks that trade-off.
The standard model gives you speed and cost efficiency. Extended Thinking gives you reasoning without the awkward pause. Together, they cover the full range of voice agent use cases.
For developers building voice agents, this is a meaningful upgrade. For businesses deploying customer service bots, it means better experiences at lower cost. For regular users, it means voice AI that finally feels like a conversation, not a command line.
The AI race isn't just about who has the smartest model. It's about who can make that intelligence feel natural. Google just made a strong case that it's winning that particular battle.
FAQ Section
What is Gemini 3.8 Live?
Gemini 3.8 Live is Google's new real-time voice model. It handles audio-to-audio conversations with visual input, automatic language switching across 97 languages, and background tool execution.
What's the difference between Gemini 3.8 Live and Extended Thinking?
Gemini 3.8 Live is built for scale and cost efficiency. Extended Thinking is built for high-complexity, multi-step tasks. Extended Thinking reasons while it speaks, narrating its progress instead of going silent.
How much does Gemini 3.8 Live cost?
Audio input costs $0.005 per minute and audio output costs $0.018 per minute through the Live API. This works out to about $1.38 to $1.50 per hour of conversation.
How much does Gemini 3.8 Live cost?
Audio input costs $0.005 per minute and audio output costs $0.018 per minute through the Live API. That's roughly $1.38 to $1.50 per hour of conversation.
How does it compare to GPT-Live-1?
On the Speech to Speech Quality Index, Extended Thinking scored higher than GPT-Live-1 Astra. On cost, Extended Thinking is cheaper at $3.50 per hour versus Astra's $5.83.
What is the Speech to Speech Quality Index?
It's a benchmark from Artificial Analysis that combines voice reasoning, agentic performance, arena preference, and task success rate into a single score.
How many languages does Gemini 3.8 Live support?
It supports 97 languages with automatic detection and mid-conversation switching.
Can I use Gemini 3.8 Live for free?
Regular users can access Gemini 3.8 Live through Search Live. Extended Thinking is available in Gemini Live and, for Pro and Ultra subscribers, in Google Docs, Gmail, and Keep.
What is SynthID?
SynthID is Google's invisible watermarking technology. All audio generated by these models is watermarked to help detect AI-generated content.
What is the asynchronous reasoning protocol?
It is the way Extended Thinking continues the conversation even while thinking. It does not pause for thinking; rather, it offers hints of what it is thinking through in performing multi-step tasks
Can developers find any chance to leverage these models?
Absolutely. These models can be accessed through the Gemini API and Google AI Studio with integration to LiveKit, Pipecat, and Agora.
