Google Releases Gemini 3.8 Live Voice Models

The division of labor is clear in the announcement. Gemini 3.8 Live combines conversational intelligence, fluid dialogue and visual grounding, tuned for high-volume deployment and cost efficiency, while Gemini 3.8 Live Extended Thinking targets high-complexity tasks with more intelligence and multi-step reasoning. For developers and enterprises they are building blocks for production voice agents, and in the Gemini app, Google Workspace and Search they become the place where you hand complex tasks over by voice.

The numbers cluster around Extended Thinking. According to the announcement, the model took the number one overall spot on Artificial Analysis' Speech to Speech Quality Index with 82.6, scored 68.6% on the agentic task-completion test τ-Voice and 35.1% on Sierra's τ-Voice-banking. Its score on the reasoning test Big Bench Audio is 97.7%. Gemini 3.8 Live, the company wrote, secured second place in the Speech Agent Arena, which measures user preference. On ServiceNow's voice-agent benchmark EVA-Bench, Google said the models push the Pareto frontier for complex workflows by balancing accuracy with conversational quality.

Seen through an engineer's eyes, what matters more than benchmarks is how the models do work in the middle of a conversation. Gemini 3.8 Live executes tools and API calls in the background while it keeps talking. It acknowledges a request and does not stop the conversation until the task is done, which removes the most awkward stretch of silence in voice agents at the design stage. It also processes visual input in near real time as context for the conversation, and it automatically detects and switches between 97 supported languages mid-conversation.

Extended Thinking adds the ability to think and talk at the same time. The announcement says the model reasons and speaks simultaneously, acknowledges prompts with short verbal cues such as "Let me check that…" and narrates the progress of multi-step background tasks as they run. Rather than eliminating wait time, it fills the wait with conversation. The problem of a user staring at a blank screen while a reasoning model composes an answer in text chat gets solved differently in voice.

The skeleton of the model is in the model card. According to the six-page Gemini 3.8 Audio model card METAL reviewed, both models are built on Gemini 3 Pro, take audio, images, video and text with a context window of up to 128K tokens, and output audio and text of up to 64K tokens. The knowledge cutoff is January 2025. For its frontier safety assessment the company used Gemini 3.7 Flash as the reference, and it wrote that the two new models show no meaningful new capabilities or material increases in performance compared with 3.7 Flash, so they are not likely to reach any tracked capability levels. METAL previously reported on the release of Gemini 3.7 Flash.

The distribution path is widest on the developer side. According to the announcement, developers can use the Gemini Live API through platforms such as Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents, and the company explained that these platforms handle the real-time media streaming infrastructure behind the scenes so developers can focus on the user experience. Salesforce, Genspark and Lumeris are working with the two models over their latency, fluidity and tool-calling capabilities. Enterprises get them first in a private preview in Gemini Enterprise, with Gemini Enterprise for Customer Experience and Google Workspace business customers coming soon.

Search is where general users meet them first. According to reports, Rajan Patel, VP of engineering for Search at Google, wrote, "New Gemini audio models just dropped – 3.8 Live is now powering real-time conversations in Search Live," adding that this brings web links attached to answers for going deeper, the ability to switch languages mid-conversation and more natural interactions. In the Google app you tap the Live icon and ask out loud, an audio answer comes back, and a transcript button shows the text or you can pick up where you left off in AI Mode history. Extended Thinking goes into Gemini Live, into Docs for Google AI Pro and Ultra subscribers, and into Gmail and Keep for all AI subscribers.

The competitive picture turns on a difference in voice-model design. METAL previously reported that OpenAI released GPT-Live-1, a full-duplex voice model that listens while it speaks, through its API. According to reports, Gemini 3.8 Live Extended Thinking finished ahead of OpenAI's latest GPT-Live-1 on the Speech to Speech leaderboard, but the rival model's full-duplex architecture, which handles listening and speaking at the same time, may still lead on conversational naturalness, and judging by the demos it sounds better too. That report concluded that Google once again put its weight on price over quality.

Every generated sound carries a mark. The company said all audio generated by its AI products is watermarked with SynthID. The imperceptible watermark is woven directly into the audio output so AI-generated content remains detectable, and the more voice agents move into phone calls and customer service, the more that mark will be used.

For teams building voice agents, the two things to take from this launch are design choices, not benchmark rankings. Tool calls do not block the conversation, and reasoning shows up as speech rather than silence. How those two hold up in real services is something the teams putting them on the Gemini API will learn first. Since the model card states there are no new capabilities compared with 3.7 Flash, the value of this release lies not in a leap in intelligence but in a change in how the model works while it talks.