Google just shipped two voice models that talk and reason at the same time. Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking arrived on September 15, 2026, and Google DeepMind calls them the company’s best conversational AI to date. Logan Kilpatrick, Google’s developer platform lead, described them as SOTA live audio models with frontier price and performance — available from day one, not behind a waitlist.
TL;DR: Google launched Gemini 3.8 Live and 3.8 Live Extended Thinking on September 15, 2026, describing them as its best conversational AI. The models handle real-time voice, support 97 languages with mid-conversation switching, and run tools in the background — topping the Artificial Analysis Speech to Speech Index at 82.6.
What Are Gemini 3.8 Live and 3.8 Live Extended Thinking?
Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are two speech-to-speech audio models introduced by Google on September 15, 2026. Google DeepMind describes them as “our best conversational AI” — models that talk, think, and handle tasks in the background without breaking your flow. In practical terms, they take audio in and produce audio out, rather than chaining a transcription model, a text model, and a voice synthesizer together.
The lineup splits into two variants. Gemini 3.8 Live is the standard real-time conversation model, built for low-latency voice interaction. Gemini 3.8 Live Extended Thinking adds near-time reasoning, simultaneous speech-and-thought processing, and background tool execution on top of that foundation.
Both variants went live at launch, available through Google’s developer platform. The names actually leaked ahead of the announcement, surfacing on Google Cloud’s quota page before the official reveal — a detail tracked by outlets like Progressive Robot.
The key specification: 97 supported languages with the ability to switch between them mid-conversation, plus asynchronous tool calls. Logan Kilpatrick confirmed these features directly in his launch announcement on X.
Why Does Speaking and Thinking at the Same Time Matter?
Because until now, voice assistants did things in sequence: they heard you, paused to generate a response, then spoke. Gemini 3.8 Live Extended Thinking collapses that pipeline. The model processes your speech and reasons over it simultaneously, which Google describes as “simultaneous speech-and-thought processing.”
What does that mean for an actual conversation? No awkward silence while the model churns. The Extended Thinking variant performs reasoning in near-time — close enough to instant that the conversation keeps its natural rhythm — and executes tools in the background. When it needs to look something up or run a function, it does so without putting the call on hold.
The 1950.ai analysis of the launch frames this as a shift in how conversational AI gets built: models that can hold a spoken conversation while quietly doing agentic work underneath. Voice stops being a thin wrapper around a text model and becomes the primary interface.
This matters for real-world use. Phone support, live translation, hands-free coding assistance, accessibility tools — all of them degrade badly when the system takes five seconds to answer. Simultaneous processing is the difference between an assistant that feels like a phone tree and one that feels like a person.
How Does Extended Thinking Work in a Voice Conversation?
Google describes three capabilities in the Extended Thinking variant: near-time reasoning, simultaneous speech-and-thought processing, and background tool execution. Each addresses a different failure point of traditional voice AI.
Near-time reasoning lets the model think through harder questions before speaking — but within the latency budget of a live call, not after a long pause. Simultaneous speech-and-thought means the reasoning happens in parallel with the conversation instead of blocking it.
Background tool execution is the most agentic piece. According to Google DeepMind’s announcement, the models “handle tasks in the background without breaking your flow.” You can ask the model to check something, book something, or fetch data mid-conversation, and it runs those calls asynchronously while the dialogue continues. Logan Kilpatrick highlighted this async tool-call support as a headline feature of the launch.
The result is a single model doing what used to require an orchestrated stack — transcription, reasoning, tool use, and speech synthesis — with the latency those separate systems imposed. The comparison points here are OpenAI’s and Anthropic’s voice modes, which both launched real-time voice experiences before Google; the aivy news coverage positions Gemini 3.8 Live as Google’s answer to those.
How Many Languages Does Gemini 3.8 Live Support?
97 languages — with seamless switching between them inside a single conversation. That figure comes straight from Logan Kilpatrick’s launch post, and it’s one of the most concrete numbers Google shared. Aivy’s coverage emphasizes the same thing: the models “handle 97 languages in one call,” meaning you don’t pick a language at the start and stay locked in.
Language switching mid-sentence is harder than it sounds. Most voice systems require you to configure a locale up front, and mixing languages breaks them. Google’s phrasing — “can seamlessly switch” — suggests the model detects and follows the language you’re actually speaking, in real time, without reconfiguration.
For developers, this collapses multilingual voice app architecture into a single model call. No language detection layer, no per-locale model routing, no separate voice assets for each market.
The 82.6 score on the Artificial Analysis Speech to Speech Index puts both models at the top of that leaderboard, per Tech-Insider’s launch coverage — meaning the multilingual capability doesn’t come at the cost of measured speech quality.
What Can Developers Build With the New Audio Models?
Google published a dedicated developer post: “Build real-time voice applications with Gemini 3.8 Live and 3.5 Transcribe.” That post pairs the new Live models with a separate Gemini 3.5 Transcribe model, giving developers a two-piece audio toolkit — one for real-time conversation, one for transcription.
The launch is aimed squarely at real-time voice workflows. DEV Community’s coverage describes the release as “Gemini 3.8 Live audio models for real-time voice AI workflows,” pointing at the practical building blocks developers get:
- Real-time speech-to-speech conversation in a single model call
- 97 languages with mid-conversation switching
- Asynchronous tool calls during live audio sessions
- Background task execution that doesn’t interrupt the dialogue
- Near-time reasoning in the Extended Thinking variant
- A companion transcription model, Gemini 3.5 Transcribe, for speech-to-text pipelines
- Frontier pricing plus performance, per the launch announcement
- Availability through Google’s developer platform at launch, not behind a waitlist
The typical use cases fall into a few buckets. Voice agents for customer support, where background tool calls handle account lookups while the conversation continues. Live interpretation apps, where the 97-language support does the heavy lifting. Voice-first productivity tools and coding assistants. And agentic applications where the model isn’t just chatting — it’s acting, fetching, and executing while it talks.
Google’s framing of “talk, think, and handle tasks” is essentially the agent loop rendered in audio. The models are available now, which makes this one of the fastest paths from announcement to buildable product in the live audio space.
Frequently Asked Questions
When were Gemini 3.8 Live and 3.8 Live Extended Thinking released?
Google launched both models on September 15, 2026. Logan Kilpatrick confirmed the release on X, noting the models are “available with frontier price + performance” — meaning they went to developers at launch rather than entering a restricted preview.
What score did Gemini 3.8 Live achieve on speech benchmarks?
The models topped the Artificial Analysis Speech to Speech Index with a score of 82.6, according to Tech-Insider’s launch coverage. That placement is what supports Google’s claim of “our best conversational AI” and SOTA status among live audio models.
What is the difference between Gemini 3.8 Live and the Extended Thinking version?
Gemini 3.8 Live is the standard real-time voice model. The Extended Thinking variant adds three capabilities on top: near-time reasoning, simultaneous speech-and-thought processing, and background tool execution — per Google DeepMind’s launch announcement and Neowin’s coverage of the release.
How many languages does Gemini 3.8 Live support?
97 languages, with the ability to switch between them seamlessly in a single conversation. Logan Kilpatrick specified this in his launch post, and aivy’s coverage confirms the models “handle 97 languages in one call” without requiring a language to be fixed at session start.
Summary
Gemini 3.8 Live and 3.8 Live Extended Thinking mark Google’s most serious move into real-time voice AI. Key takeaways:
- Two speech-to-speech models launched September 15, 2026, positioned by Google DeepMind as its best conversational AI
- The Extended Thinking variant reasons and talks simultaneously, with tool calls running in the background
- 97 languages are supported, with seamless mid-conversation switching
- The models top the Artificial Analysis Speech to Speech Index at 82.6
- A companion Gemini 3.5 Transcribe model and dedicated developer docs shipped the same day
If you build voice applications, the developer post on real-time voice apps is the best starting point. Google’s competitive position here now depends on what developers actually ship with these APIs — and how OpenAI and Anthropic respond.
How Does Gemini 3.8 Live Perform Against Competing Voice Models?
Gemini 3.8 Live and 3.8 Live Extended Thinking top the Artificial Analysis Speech to Speech Index with a score of 82.6, making them the highest-ranked live audio models at launch (Artificial Analysis, via Tech-Insider, September 2026). Google describes both as its new state-of-the-art (SOTA) live audio models, and Logan Kilpatrick highlighted “frontier price + performance” in the launch announcement on X. That combination matters because voice models historically forced a trade-off between latency and reasoning depth.
The benchmark win is only half the story. The models also handle 97 languages within a single call, switching between them without dropping the session. Competing voice products typically require language selection upfront or degrade when users code-switch mid-sentence. According to aivy.com.au, the models “top the speech benchmarks and handle 97 languages in one call,” which is a practical differentiator for global deployments.
Tom Kydd’s analysis on 1950.ai frames the launch as a shift in conversational AI architecture: the models talk, think, and execute tasks concurrently. If the benchmark results hold up under independent replication, developers building voice agents now have a clear default starting point. That was not true six months ago.
Where Are the Models Available and What Do They Cost?
Both models are available through Google’s developer platforms, with the Gemini API and Google Cloud integration confirmed in the launch documentation (blog.google, September 2026). Interestingly, the model names surfaced on Google’s Cloud quota page before the official announcement, as documented by progressiverobot.com on September 15, 2026 — an early sign that infrastructure was already provisioned for high-volume production traffic. Google positions the pair as offering “frontier price + performance,” per Logan Kilpatrick’s announcement, though exact per-token audio pricing was not detailed in the sources reviewed.
For developers, the entry point is the “Build real-time voice applications” documentation, which pairs Gemini 3.8 Live with Gemini 3.5 Transcribe. That pairing suggests Google expects developers to mix streaming conversation with high-accuracy batch transcription in the same pipeline. Neowin confirms the models are “available” broadly at launch rather than restricted to a waitlist. Consumer availability through Gemini apps was not specified in the sources, so developers building on the API are the clear first audience.
What Background Tool Execution Means for Voice Agents?
Background tool execution is the headline capability of 3.8 Live Extended Thinking. The model can call tools asynchronously while the conversation continues — the voice keeps flowing while lookups, API calls, or computations run out of band. Google DeepMind’s launch thread puts it simply: the models “talk, think, and handle tasks in the background without breaking your flow.”
Why does this matter? Traditional voice agents block on tool calls, producing awkward silence while a database query or booking request completes. With asynchronous tool calls, confirmed by Kilpatrick on X, the model can acknowledge a request conversationally, execute the tool in the background, and weave the result into the dialogue seconds later. The Neowin coverage describes this as “simultaneous speech-and-thought processing” plus “background tool execution” as distinct capabilities of the Extended Thinking variant.
For agent builders, this changes the interaction pattern fundamentally. A travel booking agent can search flights, check prices, and hold options while the user keeps talking about dates. The 1950.ai analysis argues this architecture is what powers “the next wave of AI agents” — voice-first systems that act, not merely respond.
How Does 3.8 Live Compare to ChatGPT and Claude Voice Modes?
Google positions Gemini 3.8 Live as competing directly with OpenAI’s and Anthropic’s voice offerings — aivy.com.au’s headline says it outright: “After ChatGPT and Claude comes Gemini 3.8 Live.” The distinguishing claim is benchmark leadership: the 82.6 score on the Artificial Analysis Speech to Speech Index puts it at the top of that index at launch (Tech-Insider, September 2026).
The bigger architectural difference is Extended Thinking. ChatGPT’s voice mode and Claude’s voice capabilities operate as speech interfaces over a text model; deep reasoning and low-latency conversation have remained separate modes. Gemini 3.8 Live Extended Thinking merges them, running reasoning and background tool execution inside a live audio session. Add the 97-language single-call support and asynchronous tool calls, and the feature set is broader than what either competitor documents today.
In my opinion, the honest caveat is that benchmark leadership in speech-to-speech quality does not automatically translate to everyday preference — voice UX depends heavily on app integration, latency in real networks, and accent handling. Independent, long-term testing will tell.
Are There Any Caveats With the Launch?
Yes, and they are worth planning around even if the launch itself looks strong. First, the models leaked early through Google’s Cloud quota page before official announcement, which means some developer tooling and documentation may still be settling (progressiverobot.com, September 15, 2026). Second, detailed audio pricing per minute or per token was not included in the source material, so cost planning for production workloads requires checking current Google Cloud pricing directly.
Third, Extended Thinking’s background tool execution raises evaluation questions. When a model acts asynchronously during a conversation, debugging failures and auditing what ran becomes harder than with request-response patterns. Developers will need new observability practices around tool traces. Fourth, the progressiverobot.com piece couples the launch with market forecasts about competing models like Grok 4.7, a reminder that the speech AI leaderboard is volatile — the 82.6 Artificial Analysis score is a snapshot, not a permanent ranking.
None of these caveats undermine the core launch. They simply mean production teams should validate pricing, latency, and tool-call behavior against their own workloads before committing.
Frequently Asked Questions
What is the difference between Gemini 3.8 Live and 3.8 Live Extended Thinking?
Gemini 3.8 Live is the standard state-of-the-art live audio model supporting 97 languages and asynchronous tool calls. 3.8 Live Extended Thinking adds near-time reasoning, simultaneous speech-and-thought processing, and background tool execution, letting the model run tools while the conversation continues (Neowin; Google DeepMind, September 2026).
How many languages does Gemini 3.8 Live support?
Gemini 3.8 Live supports 97 languages and can switch between them seamlessly within a single call, according to Logan Kilpatrick’s launch announcement on X. This single-session multilinguality is one of the model’s headline differentiators versus competing voice products.
Can Gemini 3.8 Live call tools during a conversation?
Yes. The models support asynchronous tool calls, and 3.8 Live Extended Thinking specifically executes tools in the background without interrupting the spoken conversation. Google DeepMind describes this as handling “tasks in the background without breaking your flow” in its launch thread.
Where can developers access Gemini 3.8 Live?
Developers can access the models through the Gemini API and Google Cloud, with Google publishing dedicated guidance titled “Build real-time voice applications with Gemini 3.8 Live and 3.5 Transcribe.” The models launched broadly on September 15, 2026, after the names briefly appeared on Google’s Cloud quota page.
Summary
Google’s Gemini 3.8 Live launch is a substantial step for real-time voice AI. Here are the key takeaways:
- Two models launched on September 15, 2026: Gemini 3.8 Live and 3.8 Live Extended Thinking, described by Google as its best conversational AI to date.
- Gemini 3.8 Live topped the Artificial Analysis Speech to Speech Index with a score of 82.6 and supports 97 languages with seamless switching in one call.
- The Extended Thinking variant introduces background tool execution and simultaneous speech-and-thought processing, so voice agents can act while the user keeps talking.
- Developers get a documented path via the Gemini API and Google Cloud, paired with the Gemini 3.5 Transcribe model for transcription workflows.
- Pricing details and production-readiness caveats mean teams should validate cost, latency, and tool-call traces on their own workloads before scaling.
If you build voice applications, the Gemini 3.8 Live documentation on Google’s developer blog is the best starting point. Test the background tool execution pattern first — it is the capability most likely to reshape how voice agents are designed.