Google released two new audio-focused Gemini models on Tuesday, targeting the fast-growing market for voice agents and real-time conversational AI. Gemini 3.8 Live and 3.8 Live Extended Thinking represent the company's bet that the next major interface for AI will be spoken, not typed.
Why voice-first AI matters now
Voice agents have been stuck in a awkward middle ground. Early systems like Siri and Alexa handled simple commands well but collapsed under complex, multi-step requests. Newer LLM-based agents can reason through hard problems but often feel clunky when forced to speak their answers. The industry needed models built from the ground up for spoken interaction, ones that could think, talk, and act simultaneously without forcing users to wait in silence.
Google is positioning these two models as a split strategy: one optimized for volume and cost, the other for raw intelligence on hard problems.
Two models, different tradeoffs
Gemini 3.8 Live is the workhorse. It is designed for production voice systems that need to handle large volumes of conversations cheaply. Google claims it scored second place in the Speech Agent Arena, a public benchmark where users rate voice agents head-to-head. The model supports 97 languages, detects visual input in near real-time, and can fire off API calls in the background while continuing to speak, so users do not sit through dead air while a tool runs.
Gemini 3.8 Live Extended Thinking is the premium option. It took the top spot on the Artificial Analysis Speech to Speech Quality Index with a score of 82.6. On agentic benchmarks, it hit 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking benchmark, both industry-leading numbers for task completion through spoken interaction. It also scored 97.7% on Big Bench Audio, a test of reasoning through audio inputs.
The key difference: Extended Thinking reasons out loud. When it needs to run a multi-step process, it gives verbal cues like "Let me check that..." and narrates its progress live. Standard 3.8 Live handles simpler tasks faster and cheaper but does not provide that running commentary.
The real-world performance picture
On ServiceNow's EVA-Bench, which evaluates voice agents on enterprise workflows, both models push what Google calls the Pareto Frontier, balancing accuracy against conversational fluency. The benchmark specifically tests how well agents handle complex, multi-turn business tasks through speech, not just simple question-and-answer exchanges.
The tool-calling distinction matters for developers building production systems. Previous voice models often forced a choice: respond quickly with a shallow answer, or pause for seconds while the model runs tools and returns a deeper response. Both 3.8 models execute tool calls and API requests in the background while continuing the conversation, which means the agent can say "I'm looking that up for you" and keep talking while the actual work happens behind the scenes.
The developer ecosystem Google is betting on
Google is not building these models in isolation. The Gemini Live API is already integrated with developer platforms including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents. These platforms handle the real-time media streaming infrastructure that makes voice agents possible, so developers can focus on conversation design rather than audio codecs and WebRTC plumbing.
On the enterprise side, Google announced partnerships with Salesforce, Genspark, and Lumeris. The company highlighted the models' latency, fluidity, and tool-calling capabilities as reasons these partners signed on.
For developers who want to experiment immediately, both models are available today in the Gemini API and Google AI Studio. Enterprise customers get private preview access through Gemini Enterprise, with broader rollout planned for Customer Experience and Workspace integrations.
What the rollout actually means
The availability timeline tells its own story. 3.8 Live is rolling out to Search Live and the Gemini app first, putting it in front of millions of users immediately. Extended Thinking follows a more measured path, targeting developers and enterprise customers before reaching consumer products like Gmail, Docs, and Keep.
All audio output from these models carries SynthID watermarks, Google's imperceptible digital fingerprint designed to identify AI-generated speech. The watermarking is baked into the audio stream itself, not appended as metadata, making it harder to strip out.
The models land at a moment when voice agent platforms are growing rapidly but struggling with the core tension between speed and depth. Google's two-model approach sidesteps that tradeoff by giving developers a choice. For high-volume customer service, use the cheaper model. For complex enterprise workflows that demand real reasoning, use Extended Thinking and let the model narrate its work instead of leaving the user in silence.