Google has launched two new voice AI models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, now available through the Gemini API, Google Workspace, and the Gemini app. Both models handle real-time reasoning, background task execution, and multilingual conversation simultaneously, without pausing the dialogue to think. For businesses building on voice interfaces, the benchmarks and architecture represent a meaningful shift in what production-ready voice agents can do.
What are Gemini 3.8 Live and 3.8 Live Extended Thinking?
The two models serve different use cases, according to Google DeepMind. Gemini 3.8 Live is built for scale and cost efficiency, combining fluid conversation with real-time visual grounding (the ability to anchor responses to what the model currently sees through a camera or screen). Gemini 3.8 Live Extended Thinking is designed for high-complexity tasks that require multi-step reasoning, such as working through a complicated support issue or a multi-turn planning task, while keeping the conversation going at the same time.
Both models support natural interruptions, multiple languages, and the ability to explain their reasoning as they work rather than disappearing and returning a single answer.
How do the benchmark numbers compare?
Google DeepMind published the following results for Gemini 3.8 Live Extended Thinking at launch in 2026:
- 82.6 on Artificial Analysis’ Speech to Speech Quality Index, ranking #1 overall across all evaluated models
- 68.6% on τ-Voice (tau-Voice), an agentic task completion benchmark that tests whether a voice AI can complete real-world tasks autonomously from start to finish, not just answer a question
- 35.1% on Sierra’s τ-Voice-banking benchmark, a specialized evaluation for voice agents handling banking-style customer support tasks
- 97.7% on Big Bench Audio, a broad audio reasoning evaluation covering comprehension, logic, and language
Gemini 3.8 Live, the more cost-efficient sibling, secured second place on the Speech Agent Arena, a community-run leaderboard where real users compare voice AI models in live conversations.
What can businesses actually build with these models?
Beyond developer toolkits, Google is also positioning Gemini Enterprise for Customer Experience (CX) as a ready-to-deploy voice agent solution aimed at retailers and restaurants, according to cloud.google.com. It uses native audio models for real-time streaming so agents can hold fluid, multi-language conversations across the full customer lifecycle: product discovery, cart building, loyalty programs, and post-purchase support.
Gemini Enterprise for CX includes a pre-built multimodal Shopping agent capable of navigating product catalogs and guiding customers through checkout. A separate Customer Experience Agent Studio lets businesses build and deploy custom support agents that connect to the Shopping agent, drawing on each customer’s historical context. A new Universal Commerce Protocol (UCP), a shared standard for commerce actions across different touchpoints, is designed so the same agent that helps a customer browse can also handle a return claim without losing context.
Does the Extended Thinking model reason out loud?
According to Google DeepMind, yes. Gemini 3.8 Live Extended Thinking is built to surface its reasoning while still holding the conversation. Background tasks, such as searching for information, pulling data, or calling external tools, run in parallel with the dialogue rather than pausing it. A user asking a complex question will hear the model working through the problem rather than waiting in silence for a completed answer.
What this means for AI-search visibility
Here is the practical shift in plain terms: in a standard web search, a brand ranked fourth still gets seen. In a voice interaction, the model typically surfaces one or two options. Ranking third often means not being mentioned at all.
That dynamic gets sharper when the model is completing tasks rather than just answering questions. A voice agent that books a table, resolves a support ticket, or builds a shopping cart autonomously removes the user from the selection process entirely. The agent becomes the decision-maker, and whichever brand the agent defaults to is the one that wins the interaction. No click, no scroll, no second chance.
The τ-Voice score of 68.6% (per Google DeepMind, 2026) is the number worth paying attention to here. It does not measure how well Gemini answers a trivia question. It measures how often a voice agent can complete a real task end-to-end without failing or handing off. That is exactly the capability that determines whether a brand appears in an agentic workflow or gets bypassed. Agents operating at this level are increasingly selecting vendors, not just surfacing results.
Hingewise’s view: the gap between “appearing in search” and “being selected by an agent” is not a future problem. Models now completing 68% of agentic voice tasks autonomously are already making selection decisions at scale. The signals agents use are different from the signals that drive traditional rankings: structured product data, API accessibility, clear machine-readable descriptions of capabilities, and up-to-date knowledge sources. Brands that have optimized for text search but not for agent-readable formats are likely already losing ground in voice contexts, even if their web traffic metrics have not shown it yet.
Before your brand deploys or audits voice AI: a quick checklist
- Test how your brand appears right now when a voice agent is asked a question in your category. Use Gemini voice, ChatGPT voice, and Perplexity voice mode directly.
- Check whether your product or service data is structured in a format AI agents can parse: schema markup, product feeds, clear FAQ pages with specific answers.
- Verify that your Google Business Profile, product catalog, and support documentation are current. Agents retrieve from recent, accurate sources.
- Confirm that your website allows Google’s agent crawlers to access key pages. Blocking crawlers blocks retrieval.
- Map the tasks a customer would ask a voice agent to do in your category, then ask: which step of that task could your brand own?
- Audit your multilingual content. Gemini 3.8 Live’s multi-language support means non-English-speaking customers can now be served at near-native quality. If your structured data is English-only, that is a gap.
What to watch next
Gemini 3.8 Live is being rolled out across Google Workspace and Search, meaning these models will increasingly shape how users interact with Google’s own products, not just third-party apps built on the API. The Universal Commerce Protocol signals Google’s intent to standardize how agents operate across commerce contexts. How broadly UCP is adopted outside Google’s own ecosystem, and whether competing platforms build compatible standards, is one of the more significant open questions to track over the coming months.
Lam Nguyen · Hingewise
Sources:
Google DeepMind, “Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking,” 2026 (deepmind.google)
Google Cloud, “Gemini Enterprise for Customer Experience,” 2026 (cloud.google.com)
Sources
This article was drafted with AI assistance from the sources cited above and checked against them before publishing. The "What this means for AI-search visibility" section is Hingewise's editorial analysis, built on the frameworks and positions we use in our GEO work. How we work: Editorial policy.