Gemini 3.8 Live Review: Google Takes #1 on Speech-to-Speech Quality With 82.6 Score

Google Just Took the Top Spot in Voice-AI Quality

Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, and the headline number is hard to ignore: Extended Thinking took #1 on Artificial Analysis' Speech-to-Speech Quality Index at 82.6 — the highest score yet on that benchmark for any model, Google's or anyone else's.

The two models are Google's most advanced live-dialogue systems yet, built for voice agents and natural spoken interaction. They are available today through the Gemini API, Google AI Studio, Google Workspace, and the Gemini app.

Two Models, Two Jobs

Google is shipping two variants with distinct positioning:

Gemini 3.8 Live is the scale-and-efficiency option. It combines conversational intelligence with fluid dialogue and visual grounding. Google says it secured second place in the Speech Agent Arena — the crowdsourced preference benchmark — while remaining "highly cost-effective" for developers and enterprises.

Gemini 3.8 Live Extended Thinking is the high-complexity option. It reasons and speaks simultaneously, using early verbal cues like "Let me check that..." to acknowledge prompts naturally while multi-step background tasks run. This is the variant that scored 82.6 on the S2S Quality Index.

The Benchmark Numbers

Here is what Google published:

• Artificial Analysis Speech-to-Speech Quality Index: 82.6 (Extended Thinking, #1 overall)
• τ-Voice (agentic task completion): 68.6%
• Sierra's τ-Voice-banking benchmark: 35.1%
• Big Bench Audio (reasoning): 97.7%
• Speech Agent Arena: #2 (Gemini 3.8 Live)
• ServiceNow EVA-Bench: Google says its models "push the Pareto Frontier for complex workflows by successfully balancing accuracy with conversational quality" — run on the Live API on Gemini Enterprise Agent Platform

The Sierra τ-Voice-banking number deserves a note: 35.1% on a banking-task benchmark is not a killer number in absolute terms, but it is a competitive one for a voice-native agent, and it is the kind of real-workload figure that matters more than synthetic evals for anyone actually building agentic voice products.

What the Models Can Actually Do

Google's blog post runs through a set of capabilities that collectively describe a voice agent that can do real work while you keep talking:

• Near-real-time visual context. Gemini 3.8 Live processes visual inputs in near real-time, so a user can show the model something and get a spoken response grounded in what it sees — the demo shows an onboarding agent answering live questions using visual context.

• 97 supported languages with mid-conversation switching. The model automatically detects and transitions between languages mid-conversation. That is a non-trivial engineering problem — language ID in a live audio stream, then a clean switch without dropping the conversation state.

• Background tool execution. The model executes tools and API calls in the background while continuing the conversation. It can acknowledge a request and keep talking while the actual work finishes behind the scenes. The demo shows a multi-step booking workflow coordinated asynchronously without interrupting the live conversation.

• Live progress narration. Extended Thinking walks users through multi-step background tasks as they progress — verbal cues about what the model is doing, not just silence while it works.

• React component generation from sketches. One demo shows Extended Thinking turning raw sketches and voice feedback into functional React components in near-real time — a live software-building assistant that takes spoken direction and produces working UI.

The Rollout

Google is rolling the models out across four surfaces:

For developers: Gemini API and Google AI Studio, starting today.

For enterprises: Private preview in Gemini Enterprise, with Gemini Enterprise for Customer Experience coming soon. The Vertex AI studio page for multimodal live is live now.

For everyone: Search Live.

For subscribers: Gemini Live, and for Google AI Pro and Ultra subscribers in Workspace in Docs, and all Google AI subscribers in Gmail and Keep — Extended Thinking specifically.

The developer availability today is the part that matters most for the AI-industry readership. The API and AI Studio access means builders can start testing these models now, which means comparison benchmarks, agentic-workload tests, and cost-per-hour analyses will start landing in the coming weeks.

How It Fits With What Google Already Shipped

This launch follows Gemini 3.8 Flash and 3.8 Flash Cyber, which Google announced separately. The Flash line is the fast, cheap tier; the Live line is the conversational, voice-native tier. The two model families are not redundant — Flash is for high-throughput, low-latency text and code workloads; Live is for spoken, real-time, multimodal interaction.

Google also recently announced agentic video understanding with Gemini, and Gemini Omni 1.1 Flash for more controlled building. The Live models are the voice-native piece of a broader agentic push.

The Cost Question

Google's blog post emphasizes that 3.8 Live is "highly cost-effective — providing developers and enterprises with a capable and efficient model built for scale." It includes a chart on Artificial Analysis cost per hour of input audio, positioning the model against other frontier options on that axis.

The chart itself is not fully legible from the blog post alone, and Google does not publish explicit per-token pricing in the announcement. For anyone building voice agents at scale, the cost-per-hour-of-input-audio number is the one to pin down — that is the operational cost that determines whether a voice agent is economically viable. Expect third-party pricing analyses to follow.

Honest Caveat

A few things to keep in mind:

• The 82.6 S2S Quality Index score is from Artificial Analysis, a third-party benchmark, which is good — it is not a self-reported internal number. But it is a single benchmark, and voice quality is multidimensional. A high S2S score does not guarantee a great experience on your specific workload.

• The τ-Voice-banking number (35.1%) is the lowest of the published figures and deserves scrutiny. Banking tasks are hard for voice agents; 35.1% is competitive but not dominant.

• The demos are Google-produced and show the model doing well on scripted scenarios. Real-world voice agent performance on messy, multi-turn, interrupt-heavy conversations is a different test — and we will not know how it performs there until third-party testing comes in.

• Google says conversations are not retained on its servers by default (relevant to the Mozilla/Mistral Smart Window zero-retention story also running this week) — but the enterprise and Workspace rollout paths have different data-handling terms that need to be read before deploying.

What This Means for Voice Agents

The voice-agent space has been waiting for a model that can do three things at once: reason, speak, and execute tools in the background without breaking conversational flow. Gemini 3.8 Live Extended Thinking is Google's answer to that combination, and the 82.6 S2S score says it is currently the best single model on that specific quality axis.

For developers, the API availability today means the comparison test can start. The relevant questions for the next few weeks: how does 3.8 Live Extended Thinking perform on τ-Voice against the other frontier voice models on real agentic tasks? What does it cost per hour of input audio in production? How does the language-switching behave on code-switched conversations? And does the background tool execution actually hold up under load, or does the demo scenario break down when the tasks are longer and the conversation is messier?

The Numbers

The Takeaway

Google's Gemini 3.8 Live Extended Thinking is the highest-scoring model yet on Artificial Analysis' Speech-to-Speech Quality Index, at 82.6, and it is available to developers today through the Gemini API. The model's headline capability — reasoning and speaking simultaneously while executing background tools — is exactly the combination voice-agent builders have been waiting for.

The banking benchmark number (35.1% on τ-Voice-banking) is the one to watch. It is competitive, not dominant, and it is the closest of the published figures to a real-workload test. Third-party voice-agent comparisons in the next few weeks will tell you whether the S2S leaderboard position translates into a better agentic experience on actual tasks.

Sources: Google Blog, "Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking," September 15, 2026; Artificial Analysis Speech-to-Speech Quality Index; ServiceNow EVA-Bench.