Two Native Voice Models, One Architectural Break
On September 15, 2026 Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, the first entries in its live conversation line that skip the pipeline most voice assistants still use. Instead of chaining speech recognition into a language model and then into synthesis, both systems operate directly on audio in an end to end path, understanding and generating at the waveform level. Google positioned them as its most advanced conversational models to date, with the central claim being parallel reasoning rather than a bigger parameter count.
A Product Split, Not a Size Split
The two models are not a small and large cup of the same thing. Standard Live optimizes for scale and cost efficiency, combining conversational intelligence, smooth turn taking and visual understanding for high volume deployment. Extended Thinking optimizes for difficult jobs where the assistant must plan, diagnose or debug across several steps. One reviewer framed the distinction as the gap between a conversational interface and a reasoning system that happens to talk.
How Reasoning While Speaking Works
Voice assistants have historically faced a bad trade. Respond instantly with shallow answers, or think properly while the caller listens to silence. Complex requests turn into hold music, which breaks the social contract of a conversation.
Extended Thinking dissolves that trade by running the reasoning chain concurrently in the background while the audible channel stays occupied. The model opens with brief spoken acknowledgements of the let me check that kind, then narrates progress as the multi step task unfolds. The user hears continuous, relevant speech instead of a gap, and the thinking happens inside the answer rather than before it.
Google chief engineer Tom Ouyang described the pair as foundation blocks for reliable voice agents that enterprises can actually put into production, which is a claim about latency and steadiness as much as about intelligence.
Benchmark Results and Their Status
Google reported figures that place the flagship at the top of several voice leaderboards. On the Artificial Analysis speech to speech quality index, Extended Thinking scored 82.6 for first place overall, ahead of OpenAI GPT-Live-1 Astra at 81.5 and Grok Voice Think Fast 2.0 at 81.3. On Big Bench Audio, which probes reasoning delivered through speech, it reached 97.7 percent. On the tau Voice agent suite it completed 68.6 percent of multi step tasks, and 35.1 percent on the banking variant.
The standard model scored 76.0 on the same quality index and placed second on the Speech Agent Arena. On ServiceNow's EVA Bench both models pushed the Pareto frontier for complex workflows, meaning accuracy rose without sacrificing conversational naturalness.
One caveat deserves emphasis: every number above comes from Google's own testing, and no independent lab had reproduced the results at publication time.
Shared Capabilities and Practical Controls
Both models accept near real time visual input, support 97 languages with automatic detection and switching mid conversation, and can fire tool and API calls in the background without interrupting the audible thread. Extended Thinking adds live progress narration around those background calls. Every clip the models generate carries Google's SynthID invisible watermark so AI audio can be detected downstream.
Per Minute Audio Pricing
The Live API bills audio by duration rather than tokens on the speech path: $0.005 per minute of input and $0.018 per minute of output, which puts a two way conversation near $1.38 per hour. Text still prices conventionally at $0.75 per million input tokens and $4.50 per million output tokens. Model IDs are gemini-3.8-live and gemini-3.8-live-extended-thinking.
Third party estimates complicate the simple version. Artificial Analysis calculated that an hour of input audio costs roughly $0.84 on standard Live against about $3.50 on Extended Thinking, a fourfold gap for a 6.6 point difference on the quality index. That ratio, not the headline score, is the actual selection question for teams building voice products.
Where the Models Land in Google's Own Products
Rollout channels differ. Standard Live is available in the Gemini API and AI Studio, in early access inside Gemini Enterprise, and to all users through Search Live. Extended Thinking covers the same developer surface, Workspace integration is listed as upcoming, and consumers meet it in Gemini Live, in Docs for AI Pro and Ultra subscribers, and in Gmail and Keep for all Google AI subscribers. Google's conversational features, the Gmail Live, Docs Live and Keep Live experiences, are built on Extended Thinking. Partner integrations were announced across Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents, with Salesforce, Genspark and Lumeris named on the enterprise side.
Why This Moves Voice Agents Forward
Until now, a voice assistant facing a hard question almost always reverted to a turn based silence, wait, deliver pattern, which is why voice felt permanently one tier below text for real work. Reasoning during speech reframes deep inference as part of the conversation instead of an interruption to it. For support lines, live coding help, interpretation and remote technical diagnosis, that is the difference between an assistant that answers and one that resolves.
Open questions remain: self reported benchmarks, materially higher runtime cost on the thinking tier, and how much background latency still leaks into conversational naturalness once traffic is real. Google's framing, that voice agents are moving from usable to production ready, will be judged against those three.
