Google unveiled Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its most advanced live-dialogue models to date: they reason while they speak, instead of thinking first and answering after. The launch landed the same day OpenAI unveiled GPT-Live, its own native voice model for ChatGPT — the race to own voice as the enterprise interface is accelerating.
Until now, most LLM-based voice assistants worked as a three-step pipeline: transcribe the audio to text, reason over that text, then synthesize the response back into speech. That process adds latency and loses nuance — tone, hesitation, emphasis — that exists only in the original audio. Gemini 3.8 Live Extended Thinking is built to reason over the audio directly and "think out loud" while it works, using natural filler phrases like "let me check that" to cover processing time without the conversation feeling like it's stalling. The model understands and generates speech in 97 languages and supports near-real-time visual grounding — meaning it can reason about what a camera sees while holding the conversation.
Google reports that Gemini 3.8 Live Extended Thinking leads Artificial Analysis's Speech-to-Speech Quality index with 82.6 points, and — more relevant for enterprise use — hits 68.6% on τ-Voice and 35.1% on τ-Voice-banking, Sierra's benchmark that measures whether a voice agent can complete real customer-service tasks in banking (verifying identity, resolving disputes, executing transactions) without human intervention. That's the number that should matter to any company evaluating whether to replace or augment a call center with AI: not how natural the model sounds, but what percentage of calls it resolves start to finish.
If your company in Guatemala or Latin America is evaluating a voice agent for customer service, collections, or first-line technical support, this launch — paired with GPT-Live on the same day — changes the selection criteria: instead of comparing only voice naturalness or cost per minute, it's worth requiring each vendor to show task-completion benchmarks in your specific domain (banking, retail, healthcare) before committing. The per-minute audio pricing ($0.005 input / $0.018 output in Gemini's case) also forces you to model the real cost of a pilot: an average 4-minute call with bidirectional audio might cost fractions of a cent, but at a scale of thousands of daily calls the math shifts fast and deserves a load test before signing an annual contract.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Get in touch for a consultation.
Contact Carlos Montiel