Google has officially released two new generation real-time voice conversation models - Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Google claims that this is its most advanced real-time conversation model to date, achieving significant upgrades in intelligence and parallel reasoning. Both models are native speech to speech models, bypassing the traditional ASR → LLM → TTS cascade architecture and directly completing understanding and generation at the audio level.
Product positioning: One product line, two missions
Google's positioning of these two models is exceptionally clear:
Gemini 3.8 Live: Designed for scale and cost efficiency, it combines conversational intelligence, smooth communication, and visual understanding capabilities, making it suitable for large-scale deployment.
Gemini 3.8 Live Extended Thinking: Designed for high complexity tasks, it has stronger intelligence and multi-step reasoning capabilities.
This is not a simple "size cup" relationship, but a fundamental fork in the product line. As OrcaRouter's review pointed out, the difference between the two lies in their different understandings of the "responsibilities of voice agents" - "one is a conversational interface, and the other is an inference system that happens to speak".
Core breakthrough: How does "thinking while speaking" work?
The fundamental contradiction of traditional voice AI
In previous mainstream voice interaction architectures, AI systems faced a dilemma: either respond quickly but only engage in shallow conversations, or engage in deep thinking but require users to endure long periods of silent waiting. After users raise questions, they often have to listen to prompts such as' thinking 'when faced with complex tasks, resulting in a fragmented and unnatural experience.
Solution to Gemini 3.8 Live Extended Thinking
Gemini 3.8 Live Extended Thinking endows AI with the ability to "Think as it Speaks" in the background for extended reasoning: the system concurrently executes complex multi-step logic disassembly, code troubleshooting, or technical diagnosis in the background while maintaining real-time coherent dialogue.
Specifically, the model synchronously outputs speech while reasoning, responding naturally to early language prompts such as "Let me check that...", and guiding users to understand the progress of multi-step backend tasks through real-time progress narration. This means that users no longer need to wait for silence in the face of complex problems - the model will process the task while informing progress in natural language, maintaining the coherence of the conversation.
Google Chief Engineer Tom Ouyang stated in his official blog that these models provide developers and businesses with the foundational modules to build reliable, production ready voice intelligent agents.
Benchmark testing: multiple evaluations topped the list
According to the test data released by Google:
| benchmarking | Gemini 3.8 Live Extended Thinking | Explanation |
| Artificial Analysis Speech-to-Speech Quality Index | 82.6 (ranked first overall) | Exceeding OpenAI GPT-Live-1 Astra (81.5%) and SpaceXAI Grok Voice Think Fast 2.0 (81.3%) |
| Big Bench Audio (reasoning ability) | 97.7% | Evaluation of Speech Reasoning Ability |
| τ - Voice (intelligent agent task completed) | 68.6% | Multi step task completion rate |
| Sierra τ-Voice-banking | 35.1% | Voice tasks in banking business scenarios |
Gemini 3.8 Live Standard Edition scored 76.0 in the Artificial Analysis Speech to Speech Index and ranked second in the Speech Agent Arena. In ServiceNow's EVA Bench voice intelligent agent benchmark test, both models have pushed up the Pareto front of complex workflows, successfully balancing accuracy and dialogue quality.
It should be pointed out that the above data were reported by Google itself, and there are currently no independent laboratory replicated results.
Comparison of Key Abilities
| capability dimension | Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking |
| Core positioning | Scale deployment and cost efficiency | High complexity tasks and multi-step reasoning |
| Reasoning mode | Quick response dialogue | Thinking while speaking, reasoning and speaking in sync |
| visual input | Near real-time processing | Near real-time processing |
| multilingual | 97 languages, automatic switching during conversations | 97 languages, automatic switching during conversations |
| Backend tool call | Support, uninterrupted conversation | Support, with real-time progress description attached |
| Artificial Analysis Index | 76.0 | 82.6 |
Developer Access and Ecology
Developers can access two models through the Gemini Live API, with model IDs gemini-3.8-live and gemini-3.8-live extended linking. Google has established integrations with multiple developer platforms, including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, Vision Agents, and more. In addition, Google has also partnered with companies such as Salesforce, Genspark, Lumeris, etc. to promote ecosystem development.
Product landing: from Search to Workspace
The push channels for the two models have some emphasis:
Gemini 3.8 Live has started pushing: developers can use it on Gemini API and Google AI Studio; Enterprise users can experience Gemini Enterprise first; All users can experience it on Search Live.
Gemini 3.8 Live Extended Thinking has also started pushing: developer channels are the same as above; Enterprise users can experience Gemini Enterprise first and will soon be logging into Google Workspace; Regular users can experience it on Gemini Live, Google AI Pro and Ultra subscribers can use it on Docs, and all Google AI subscribers can experience it on Gmail and Keep.
This push strategy is consistent with Google's recent overall layout in its product line - conversational features such as Gmail Live, Docs Live, Keep Live, etc. are all built on the Extended Thinking model.
Industry significance: Voice AI moves from "conversational" to "operational"
The release of Gemini 3.8 Live series signifies that voice AI is crossing a critical threshold. Prior to this, voice assistants were almost inevitably trapped in a turn based mode of "silence wait output" in the face of complex tasks, and there was an essential gap between user experience and text interaction. The "speak and think" mechanism of Gemini 3.8 Live Extended Thinking transforms deep reasoning from an "enemy of dialogue" to a "part of dialogue" - the model is no longer thinking before speaking, but thinking while speaking, using language itself to carry the reasoning process.
For scenarios such as intelligent customer service, real-time programming assistance, simultaneous interpretation, and remote technical diagnosis, this means that for the first time, voice agents have the ability to solve problems while chatting with human experts. Google positions this progress as a crucial step in the transition of voice intelligent agents from usable to producible.
