01
Overview
Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are Google’s newest real-time dialogue models, launched September 15, 2026. They are designed for voice-first agents that can keep a conversation moving while using tools, processing visual context and, in the Extended Thinking variant, reasoning through complex multi-step tasks in the background.
02
What is Gemini 3.8 Live?
Gemini 3.8 Live is the scale- and cost-oriented model in Google’s new Live family. Google positions it for fluid dialogue, conversational intelligence and near-real-time visual grounding. It can execute tools and API calls asynchronously while continuing to stream audio, which matters for voice agents that would otherwise pause awkwardly while waiting for a backend task.
Google says the model automatically detects and transitions between 97 supported languages mid-conversation. It can also use live visual context, allowing an agent to reason about what a user is seeing as well as what they are saying.
03
What is Gemini 3.8 Live Extended Thinking?
Gemini 3.8 Live Extended Thinking targets harder, multi-step workflows. It adds configurable thinking and can reason in the background while continuing the spoken interaction. Google describes behaviors such as acknowledging a request, narrating progress and continuing the conversation while tools or reasoning complete.
This is a meaningful distinction from a simple speech-to-text → text model → text-to-speech pipeline. Google is positioning the Live models as native speech-to-speech systems with tool use, multimodal context and reasoning integrated into the interaction.
04
Release date and availability
Google announced both models on September 15, 2026. Both are rolling out through the Gemini API and Google AI Studio.
Gemini 3.8 Live is also rolling out in Search Live. For enterprises it is in private preview in Gemini Enterprise and is planned for Gemini Enterprise for Customer Experience.
Gemini 3.8 Live Extended Thinking is rolling out in Gemini Live. Google also lists availability for Google AI Pro and Ultra subscribers in Workspace Docs, and for Google AI subscribers in Gmail and Keep. Enterprise availability is currently more limited, including private preview and announced coming-soon surfaces.
Availability can vary by account, region and product surface, so users should verify the current Google interface or API documentation rather than assuming every surface has identical access.
05
Gemini 3.8 Live API pricing
Google’s September 15 developer announcement lists Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking at $0.005 per minute for audio input and $0.018 per minute for audio output through the Live API.
Those are the launch figures Google published for audio. Production teams should recheck the live pricing page before deployment because pricing, quotas and billing details can change and other modalities or features may have separate accounting rules.
06
Core capabilities
Asynchronous function calling
The models can execute API and tool calls in the background while continuing to stream audio. This can reduce dead air in workflows such as bookings, support, onboarding and sales qualification.
Live visual context
The models can ground dialogue in visual inputs. Google demonstrates this for troubleshooting, live guidance and visually grounded tasks.
Multilingual dialogue
Google states support for 97+ languages with automatic language detection and accent consistency. The broader launch post describes automatic transitions between supported languages during a conversation.
Alphanumeric precision
Google highlights improved handling of confirmation codes, claim numbers and technical data—an important capability for practical voice agents where a single wrong character can break a workflow.
Incremental content updates
The developer announcement says the models can merge real-time audio with structured data to produce context-aware responses as information changes.
Configurable thinking
Extended Thinking supports configurable thinking for complex reasoning. This should be treated as a capability of the Extended Thinking variant, not assumed to behave identically in standard 3.8 Live.
07
Benchmarks: what Google reports
Google says Gemini 3.8 Live Extended Thinking took the #1 overall position on Artificial Analysis’ Speech-to-Speech Quality Index with a score of 82.6. Google also reports 68.6% on τ-Voice, 35.1% on Sierra’s τ-Voice-banking benchmark and 97.7% on Big Bench Audio.
Gemini 3.8 Live is reported as second on Speech Agent Arena. Google also says the models push the Pareto frontier on ServiceNow EVA-Bench for balancing task completion and conversational experience.
These figures should be read with attribution. Some underlying benchmarks are external, but Google selected and presented the results in its launch material. They do not guarantee performance for a particular voice-agent workload.
08
Gemini 3.8 Live vs Extended Thinking
Use Gemini 3.8 Live when conversational speed, scale and cost efficiency are the priority. Consider Extended Thinking when the agent needs deeper multi-step reasoning while preserving a natural spoken interaction.
The distinction is not simply “cheap model versus expensive model.” Google currently presents both as part of the same Live family and quotes the same launch audio input/output rates in its developer announcement. The practical difference is the deeper configurable reasoning behavior of Extended Thinking.
09
Voice-agent integrations
Google names Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents as developer platforms supporting the Gemini Live API ecosystem. These platforms can handle real-time media infrastructure so application teams can focus on agent logic and user experience.
Support in an integration does not mean every Gemini Live capability is automatically exposed. Developers should verify the specific framework’s current model support, streaming behavior and tool-calling implementation.
10
SynthID and generated audio
Google says all audio generated by its AI products is watermarked with SynthID. The watermark is designed to remain detectable and help identify AI-generated audio.
SynthID should not be described as a guarantee against misuse or as proof that every third-party recording pipeline preserves detectability under every transformation.
11
Who should evaluate Gemini 3.8 Live?
The clearest use cases include customer-service agents, conversational sales and qualification, onboarding, real-time tutoring, live troubleshooting, booking agents, voice interfaces for business software and multimodal assistants that need to see and hear the user while taking actions.
Creators and marketers may also find the model useful for voice-driven content workflows, interactive experiences, conversational research and tools that generate plans or marketing assets from spoken instructions.
12
Limitations and open questions
The Gemini API release notes mark both Live models generally available as of September 15, 2026. Other Google product surfaces can still roll out separately, and rate limits, regional access and account-level quotas should be reverified before production deployment.
Benchmark numbers are not substitutes for workload testing. Voice-agent quality depends on latency, interruption handling, tool reliability, speech recognition, synthesis quality and end-to-end task completion—not a single leaderboard score.
13
FAQ
When did Gemini 3.8 Live launch?
Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026.
Is Gemini 3.8 Live available through an API?
Yes. Google says both models are rolling out through the Gemini Live API and Google AI Studio.
How much does Gemini 3.8 Live cost?
Google’s launch-day developer announcement lists $0.005 per minute for audio input and $0.018 per minute for audio output. Recheck current pricing before production use.
What is Extended Thinking?
It is the higher-reasoning Live variant. Google says it can perform configurable, multi-step reasoning in the background while continuing to speak or narrate progress.
Can Gemini 3.8 Live use tools while talking?
Yes. Google documents asynchronous function calling, allowing tool and API calls to run while the audio conversation continues.
Does Gemini 3.8 Live support visual input?
Yes. Google says it can use near-real-time visual context to ground conversations.
How many languages does it support?
Google states 97+ languages in its developer announcement and describes automatic language switching in its launch post.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.