Google Gemini 3.8 Live and Extended Thinking: Real-Time Voice AI That Thinks While It Talks

 

Quick Takeaways

• Gemini 3.8 Live is designed for fast, natural, low-latency voice conversations with speech-to-speech interaction, visual context, and tool integration.

• Live Extended Thinking adds deeper multi-step reasoning while allowing the AI to continue interacting with users and working with external tools in the background.

• The models could support new voice-based applications in customer service, education, research, productivity, and business automation through the Gemini API and Google AI development platforms.

Voice AI is moving beyond simple question-and-answer exchanges. The next generation of assistants is expected to understand spoken requests, respond naturally, use external tools and complete multi-step tasks without making conversations feel slow or mechanical.

Google is addressing that challenge with Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two voice-focused AI models announced on 15 September 2026. Designed for real-time, bidirectional audio interaction, the models combine low-latency conversation, multimodal understanding, background tool execution and, in the Extended Thinking version, deeper reasoning during live interactions.

The central idea is straightforward: an AI assistant should not have to stop being conversational simply because it needs time to think, retrieve information or complete a task.

For developers and businesses building voice agents, this could be an important shift from assistants that merely answer questions to systems capable of helping users accomplish useful work.

What Are Gemini 3.8 Live and Live Extended Thinking?

The two models belong to Google's Gemini 3.8 family but serve different purposes.

Gemini 3.8 Live is designed for fast, natural and scalable voice conversations. It targets applications where responsiveness, smooth turn-taking and low operating costs are important.

Gemini 3.8 Live Extended Thinking is built for more complex interactions. It can reason through multi-step problems, coordinate tools and continue communicating while longer-running work takes place in the background.

Both models are designed around native audio-to-audio interaction. Rather than relying exclusively on a conventional sequence of speech recognition, text generation and text-to-speech, they are intended to create a more integrated conversational experience.

This architecture can reduce some of the delays and awkward pauses associated with traditional voice systems, although the final experience will still depend on the application, network and supporting infrastructure.

Key Features at a Glance

Both models arrive packed with capabilities aimed squarely at production-grade voice agents:

• Native audio-to-audio processing — direct speech-to-speech handling eliminates cascaded pipeline latency

• Asynchronous function calling — API and tool calls run in the background without stalling the conversation

• Visual context grounding — point your camera at a broken bike chain or a leaky pipe, and the agent guides you in near real time

• 97+ languages with live switching — the model detects and transitions between languages mid-sentence, accents intact

• Alphanumeric precision — confirmation codes, claim numbers, and serial strings, a notorious weak spot for voice AI, handled with high accuracy

• Incremental content updates — responses stream in progressively (destination confirmed → dates → prices → booking options) rather than arriving after everything finishes

• Configurable reasoning depth — Extended Thinking offers low, medium, and high thinking levels, letting developers balance speed, cost, and accuracy

• SynthID audio watermarking — all generated speech carries an invisible watermark for authenticity and safety

• Generous context windows — up to 131,072 input tokens and 65,536 output tokens for longer, richer sessions

The Benchmarks: Numbers That Turn Heads

Independent evaluations put the Extended Thinking model at or near the top of the voice AI field:

That latency figure deserves attention. Cutting first-audio response from nearly three seconds to just over one is the difference between talking at a machine and talking with one.

One honest caveat: in blind human-preference listening tests, some listeners still favoured Google's earlier, lighter Flash Live model. The new Extended Thinking model completed more tasks and scored higher on benchmarks, but raw "sounds natural" appeal remains a moving target. Benchmarks measure capability; ears measure comfort. The gap between the two is where the next round of competition will be fought.

Availability: Where You Can Actually Use It

Google isn't keeping these models locked in a lab; they're already woven across its ecosystem.

For developers:

• Available now via the Gemini API and Google AI Studio under the model IDs gemini-3.8-live and gemini-3.8-live-extended-thinking

• Streaming integration through WebSockets with support from partner platforms including LiveKit, Vercel, Agora, LangChain, Pipecat, Fishjam, and Vision Agents

For businesses:

• Private preview through Gemini Enterprise, with customer-experience-focused tooling rolling out soon

• Early adopters are already deploying the models for multilingual training simulations and autonomous support workflows

For everyday users:

• Search Live runs on the standard Gemini 3.8 Live model

• Extended Thinking is arriving in the Gemini app and Google Workspace, powering conversational features in Gmail, Docs, and Keep (Docs requires an AI Pro or Ultra subscription)

Pricing: The Not-So-Secret Weapon

Here's where Google applies real pressure. The published rates:

Translated into real-world terms: an hour of Extended Thinking voice conversation runs roughly $3.50, undercutting comparable offerings from OpenAI (~$4.47) and xAI (~$4.80). The standard model's input cost of about $0.84/hour is among the lowest on the market, period.

For startups and enterprises deploying voice agents at thousands of concurrent calls, that pricing gap isn't a rounding error. It's the difference between a viable business model and a money pit.

Getting Started: A Quick-Start Guide

For Developers

1. Open Google AI Studio and sign in with your Google account

2. Grab an API key from your dashboard and review usage limits and pricing

3. Choose your model — gemini-3.8-live for fast conversational apps, or gemini-3.8-live-extended-thinking for multi-step reasoning workflows

4. Establish a WebSocket session for bidirectional audio streaming

5. Define asynchronous tools — for Extended Thinking, function calls must be non-blocking so the model can keep talking while work runs

6. Handle the new session lifecycle — this one's critical: with Extended Thinking, a turnComplete signal no longer always means the model is idle. Monitor IN_PROGRESS and IDLE interaction states to distinguish "finished speaking" from "task actually done"

7. Test ruthlessly — interruptions, accents, background noise, code-switching, visual inputs, ambiguous requests

8. Add safety rails — logging, consent controls, human escalation paths, and verification for high-impact actions

For Everyday Users

1. Open the Google app or Gemini app and look for Live or Search Live

2. Start talking — try pointing your camera at a problem for visual help

3. Switch languages naturally mid-conversation, or interrupt whenever you like

4. Workspace subscribers can activate Extended Thinking in supported Gmail and Docs side panels

The Analysis: Why This Launch Actually Matters

It Kills a Long-Standing Trade-Off

Until now, voice AI developers faced a frustrating fork: fast models that couldn't reason, or reasoning models that couldn't hold a real-time conversation. Extended Thinking's background-processing architecture collapses that trade-off. A customer can describe a messy billing dispute and get genuine problem-solving, not a canned "let me transfer you."

That's an architectural advance, not an incremental one. It moves voice AI from the rigid request-response pattern that has defined chatbots for a decade toward agents that feel like collaborators.

Voice Is the Next Battleground, and Google Just Fired the First Cheap Shot

Voice is widely viewed as the next major interface layer: the point where AI stops being a text box and becomes a conversational partner. Whoever owns low-latency, high-intelligence, affordable speech owns a massive slice of the next decade of human-computer interaction.

OpenAI's full-duplex model still arguably edges out Google on raw conversational naturalness. But Google is playing a familiar game: compete on scale, ecosystem integration, and price while closing the quality gap release after release. With Search, Workspace, Android, and Cloud all serving as distribution channels, Google can embed this capability everywhere at once, something no rival can currently match.

The Rise of Agentic Voice

Perhaps the deeper signal is what these models represent: the shift from voice assistants that answer to voice agents that act. Retrieving order statuses, cross-referencing travel options, diagnosing technical problems, completing bookings, all while sounding like a helpful human on the other end of the line. Early enterprise deployments in finance and customer support suggest the production era of agentic voice isn't coming. It's here.

The Challenges Google Still Faces

No launch is a guaranteed win, and the hurdles are real:

• Hallucinations sound worse out loud. A confident, fluent voice delivering wrong information is more convincing, and more dangerous, than a text error

• Privacy remains unresolved. Always-listening agents processing sensitive conversations and camera feeds will face scrutiny from regulators and users alike

• Engineering complexity is up. Managing async session states, deduplicating tool calls, and handling delayed responses adds genuine production overhead

• Costs scale with conversation. Per-minute pricing looks cheap in demos but compounds fast with long calls, heavy tool use, and large user volumes

• The "naturalness" gap. Until the new models win human blind-listening tests decisively, rivals have a quality story to tell

And you can bet OpenAI, Meta, Microsoft, Amazon, and xAI won't sit still. A price war in voice AI ultimately benefits everyone building on these platforms.

Bonus: Gemini 3.5 Transcribe Rounds Out the Audio Stack

Alongside the live models, Google is also promoting Gemini 3.5 Transcribe, a dedicated speech-to-text model covering 85+ languages with automatic code-switching and custom vocabulary support. Reported word error rates sit at 4.0% for streaming and 2.6% for non-streaming transcription, solid numbers for live captions, call-centre analytics, and meeting notes. Together, the releases give developers a full audio toolkit: listening, understanding, speaking, and transcribing.

The Bottom Line

Google's Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking represent an important development in real-time voice AI. Gemini 3.8 Live focuses on fast, natural dialogue, while Live Extended Thinking is designed for more complex reasoning and background tool execution. Together, they reflect a broader movement towards AI agents that can listen, understand, reason, speak and act within the same conversation.

For developers, the tools to build genuinely useful, multilingual, action-oriented voice agents are finally within reach. For businesses, production-grade voice automation just got dramatically cheaper. For competitors, the bar didn't just move; it moved and dropped in price.

The most important question, however, is not simply whether AI can speak naturally. It is whether voice agents can consistently complete useful tasks while remaining accurate, transparent, secure and respectful of privacy. If those challenges can be addressed, voice AI may become less like a conventional chatbot and more like a genuinely useful digital collaborator

What would you build with a voice agent that can think while it talks? Drop your thoughts in the comments. 👇

Shakir Bukhari

Related Reads

• Google Gemini Adds Setting to Remove Visible Media Watermarks; AI Images, Videos and Music Can Now Go Clean /2026/08/google-gemini-adds-setting-to-remove.html

• Google Gemini Hits 1 Billion Monthly Users: Google’s Fastest-Growing Product Challenges ChatGPT /2026/08/google-gemini-hits-1-billion-monthly.html

• Google Gemini Just Got a Superpower: How Memory Turns Live Chats Into a Personal Assistant /2026/06/google-gemini-just-got-superpower-how.html