On September 15, 2026, Google released two new real-time voice models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, built for conversational speech-to-speech interaction rather than the read-a-transcript style of voice assistant most people have used until now. Both process spoken audio in and out directly, switch between 97 languages mid-conversation, and can run tool calls or API requests in the background without stopping to think out loud. Google says the models beat OpenAI's GPT-Live-1 (Astra) and xAI's Grok Voice Think Fast 2.0 on a widely cited third-party quality benchmark, at a lower price per hour.
This is a release worth slowing down on, because it says something about where real-time voice AI is heading in 2026: away from stitched-together pipelines and toward models that hear and speak natively, with a cheap fast tier and a pricier reasoning tier sitting side by side.
What Google shipped
Gemini 3.8 Live is the standard tier, tuned for low latency and fluid back-and-forth conversation. Gemini 3.8 Live Extended Thinking adds a reasoning layer on top, letting the model work through a harder problem, such as a multi-step booking or a technical support diagnosis, while still talking, using short verbal acknowledgment cues to cover the extra processing time. Google's launch post says Extended Thinking took the top spot on Artificial Analysis' Speech to Speech Quality Index, and both figures below come from that same benchmark as reported by officechai.com and Google's own announcement.
Component benchmark numbers reported alongside the headline index include 68.6% on the agentic tau-Voice benchmark and 97.7% on Big Bench Audio for Extended Thinking, plus a second-place finish for the standard tier on the separate Speech Agent Arena leaderboard. These are the benchmark provider's published scores, not results independently reproduced, and they should be read the way any vendor-cited benchmark should: directionally informative, not a guarantee of real-world performance on a specific task.
Google
OpenAI
xAISpeech-to-speech versus the old cascade
Most voice assistants people have used, from early Alexa-style devices through last year's chat apps, run what is usually called a cascade architecture: automatic speech recognition converts your voice into text, a text-based language model reads that transcript and writes a text reply, and a text-to-speech engine reads the reply back out loud. Three separate systems, each with its own processing delay, stitched together in a pipeline.
That pipeline works, but it loses information at every seam. The language model never actually hears your voice, so it cannot pick up on tone, hesitation, a rising or falling inflection, or the fact that you interrupted yourself mid-sentence. It only sees flattened text. The output side has the same problem in reverse: a text-to-speech engine reading a reply out loud has no idea what emotional weight that sentence should carry, because nothing about emotion survived being converted into words on a page.
Native speech-to-speech models like Google's Gemini 3.8 Live are built differently. Audio goes in and audio comes out of the same model, without a text bottleneck in between. That has three practical consequences worth naming directly:
- Latency drops, because there is no round trip through a separate transcription step and a separate voice synthesis step. Fewer seams means fewer places for delay to accumulate.
- Prosody and emotion carry through, because the model that decides what to say is the same model that heard how you said your part, so it can match tone rather than guessing at it from a flat transcript.
- Interruption handling gets more natural, because the model is listening continuously rather than waiting for a full utterance to be transcribed before it starts composing a reply, so it can notice when you cut in and adjust.
Gemini's Live API documentation describes this as a stateful, low-latency connection designed for real-time voice and vision, distinct from sending the model a pre-recorded audio file to process after the fact.
Speech to Speech Quality Index, by voice model
Gemini 3.8 Live Extended Thinking edges out GPT-Live-1 and Grok Voice Think Fast 2.0 on Artificial Analysis' composite score, while costing less than either per hour of audio. The standard Gemini 3.8 Live tier trades some quality for a much lower price.
Source: Artificial Analysis Speech to Speech Quality Index, as reported by officechai.com and Google's Sept 15, 2026 launch post. Third-party benchmark scores, not independently reproduced by Metir.
A cheap tier and a reasoning tier, priced apart
The two-model structure is itself a signal. Rather than shipping one voice model and calling it done, Google split the release into a fast, inexpensive default and a more expensive tier that spends extra compute reasoning through harder requests before or while it answers. At $0.84 an hour, the standard Gemini 3.8 Live tier is priced for high-volume, always-on use cases, a customer support line, an in-app assistant, a device that listens most of the day. At $3.50 an hour, Extended Thinking is priced for conversations where getting the answer right matters more than shaving off a fraction of a second, a banking dispute, a technical diagnosis, a multi-step travel rebooking.
A voice model that can reason without going silent is solving a problem cascade architectures never had to solve, because thinking and speaking were always two separate systems.
Analysis
Google's own comparison, as reported by officechai.com, puts GPT-Live-1 (Astra) at $5.83 an hour and Grok Voice Think Fast 2.0 at $4.80 an hour, with Speech-to-Speech Index scores of 81.5 and 81.3 respectively, both a shade behind Extended Thinking's 82.6. If those figures hold up under independent testing, Google's pitch is a narrow quality edge over rival flagship voice models at meaningfully lower cost per hour, with a second, much cheaper tier for workloads that do not need the reasoning overhead at all.
Why mid-conversation language switching and background tool calls matter
Two capabilities in this release matter less as benchmark trivia and more as signals of what voice agents are actually being asked to do now. The first is automatic detection and switching between 97 languages within a single conversation, without the user restarting the session or picking a language from a menu. For a support line or a travel agent handling callers who code-switch mid-sentence, that removes a hard failure mode that cascade systems have long struggled with, since a speech recognizer tuned for one language often mangles a sentence that drifts into another.
The second is background tool execution: the model can kick off an API call, a database lookup, or a function the developer has wired in, and keep talking while that call resolves, rather than going silent until the result comes back. That is a small technical detail with a large practical effect, because a voice agent that has to pause mid-conversation every time it needs to check an order status or pull a record feels broken in a way a text chatbot pausing to "think" does not. A voice channel has no equivalent of a typing indicator that reads naturally; silence just reads as a dropped call.
Where it is available, and what it is not
Both models are live now in the Gemini API and Google AI Studio for developers, in a Gemini Enterprise private preview, and in Search Live for the general public. Extended Thinking additionally reaches Google Workspace, inside Docs, Gmail, and Keep, for Pro, Ultra, and All subscribers, and inside the Gemini Live app itself. Google says partners including Salesforce, Genspark, and Lumeris have already tested the models in production voice agents.
One distinction is worth being precise about, since the two products share a version number and launched close together: Gemini 3.8 Flash, Google's standard text and multimodal model released earlier this month, does not have Live API support. It ingests audio as a file for later processing rather than holding a live, bidirectional voice session. The speech-native capability described in this piece lives entirely in the separate Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking models, not in the Flash line.
The bigger picture
What Google, OpenAI, and xAI are now competing on is not "can a model talk," which has been solved in some form for years, but how naturally it can hold a conversation, at what latency, in how many languages, and at what price per hour. That the leading three voice benchmark scores as reported here sit within about 1.3 points of each other, 82.6, 81.5, 81.3, suggests the frontier on raw conversational quality is getting crowded, and that price per hour and feature depth, like background tool calls and language switching, are becoming the real differentiators.
For teams building voice products, that argues for treating the underlying voice model as a swappable component rather than a permanent commitment, the same way many teams already treat text models. Metir AI takes that approach across chat and voice models from multiple providers in one place, so a change in the leaderboard does not mean rebuilding an integration.
Sources:
- Google releases Gemini 3.8 Live-Extended conversational model, claims better performance than GPT-Live-1, Astra and Grok Voice Think Fast 2.0 at lower price - officechai.com
- Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking - Google's official launch post
- Gemini API changelog - ai.google.dev
- Gemini Live API documentation - ai.google.dev
- Gemini 3.8 Flash: Complete Guide 2026 - codersera.com, on why the text-only Flash model has no Live API support
- AI Weekly - aiweekly.co
Image credits
Hero image: Google campus signage in Mountain View, California. Photo by brionv via Wikimedia Commons, licensed under CC BY-SA 2.0. In-body photo: Google's Bay View campus complex in Mountain View, California, opened in 2022, by Dicklyon via Wikimedia Commons, licensed under CC BY-SA 4.0.