Google DeepMind has released its newest live dialogue models: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Designed to power enterprise voice agents and process multimodal inputs in near real-time, the Extended Thinking model immediately secured the top spot on the Artificial Analysis Speech to Speech Quality Index with a score of 82.6.
According to the announcement, Google has split this release into two tiers. Gemini 3.8 Live is built for scale and cost efficiency, securing second place in the Speech Agent Arena. Meanwhile, Gemini 3.8 Live Extended Thinking is tailored for complex workflows and multi-step reasoning. The extended model scored 97.7% on Big Bench Audio and leads agentic task completion benchmarks with 68.6% on τ-Voice and 35.1% on Sierra’s τ-Voice-banking benchmark.
Both models introduce technical capabilities that push past traditional conversational AI limits. They process visual inputs in near real-time to provide contextual answers and can automatically detect and transition between 97 supported languages mid-conversation. Most notably for business automation, these models execute tools and API calls in the background. The AI can acknowledge a user's request and maintain the conversation fluidly while backend tasks finish. For tasks demanding deeper logic, the Extended Thinking model reasons and speaks simultaneously without interrupting the workflow.
What This Means for SMB Operations
For small and mid-sized businesses, this leap in voice agent capability translates directly to how administrative and customer-facing workflows can be automated. Historically, voice-based service bots have suffered from high latency and rigid dialogue trees. When a legacy bot had to fetch a customer record or trigger a workflow, it resulted in awkward silences, dropped contexts, or forced the user to wait on hold.
Gemini 3.8 Live addresses this bottleneck by decoupling the conversational layer from the task execution layer. Because the model can run API calls in the background while keeping the conversation going, a business can deploy a voice agent that behaves more like a human dispatcher. For example, a local property management firm could use a voice agent to take a tenant's maintenance call. The agent can detect if the caller is speaking one of the 97 supported languages, switch to it instantly, and trigger a work order via API in the company's CRM—all while verbally confirming the next steps with the tenant without dead air.
The two-tiered model release also gives smaller businesses control over their automation budgets. Google built the standard Gemini 3.8 Live model for scale and cost efficiency. Businesses can deploy this baseline model for high-volume, low-complexity tasks like inbound call routing, basic scheduling, and employee onboarding. When an interaction requires multi-step logic—like a customer needing to alter a complex service contract or a field worker diagnosing an error—the system can route to Gemini 3.8 Live Extended Thinking to handle the deeper reasoning.
Furthermore, the addition of real-time visual context introduces new capabilities for field operations and inventory management. A field technician or warehouse worker could point a camera at a malfunctioning part or an unlabeled pallet and talk through a troubleshooting process with the AI. The model processes the visual input in near real-time, matching what it sees with backend data to provide immediate voice-guided assistance.
These tools are available now through the Gemini API, Google Workspace, Search, and the Gemini Enterprise Agent Platform. By successfully balancing accuracy with conversational quality—a metric validated on ServiceNow’s EVA-Bench—Google is making production-ready voice automation highly accessible. For businesses looking to automate administrative tasks and front-line communications, voice AI is moving past simple transcription and becoming a reliable operational interface.