GOOGLE AI GUIDE • FACT-CHECKED SEPTEMBER 19, 2026
Gemini 3.8 Live is Google’s generally available, low-latency audio-to-audio model for real-time voice agents and live dialogue. Google released it on September 15, 2026, alongside Gemini 3.8 Live Extended Thinking, a higher-reasoning version built for complex, multi-step voice tasks. Google’s announcement was updated on September 17. [1] [2]
Quick answer: Use Gemini 3.8 Live when responsiveness, scale and natural real-time dialogue matter most. Use Gemini 3.8 Live Extended Thinking when a live voice agent needs more background reasoning for difficult workflows. For speech-to-text, Google offers the separate Gemini 3.5 Transcribe family, which supports more than 85 languages. [3]
Gemini 3.8 Live at a glance
- Release date
- September 15, 2026
- Release status
- Generally available (GA)
- Standard model ID
gemini-3.8-live- Extended model ID
gemini-3.8-live-extended-thinking- Inputs
- Text, images, audio and video
- Outputs
- Text and audio
- Input token limit
- 131,072 tokens
- Output token limit
- 65,536 tokens
Google’s API documentation describes Gemini 3.8 Live as the default option for most low-latency voice-agent experiences, while Extended Thinking is recommended when higher background reasoning is required. [4] [5]
What is Gemini 3.8 Live?
Gemini 3.8 Live is a native audio-to-audio Gemini model optimized for real-time dialogue. It is designed to receive live multimodal context, generate spoken responses and use tools without forcing every interaction through a conventional speech-to-text → text model → text-to-speech pipeline.
For users, the intended result is a more fluid conversation. For developers, the more important change is architectural: the model can combine voice, visual context, reasoning and external tool calls in one live session. [1]
Gemini 3.8 Live vs. Gemini 3.8 Live Extended Thinking
| Area | Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking |
|---|---|---|
| Best for | Low-latency voice agents and real-time dialogue | Complex, multi-step voice tasks |
| Reasoning | Interleaved reasoning optimized for responsiveness | Higher background reasoning during live interaction |
| Function calling | Supported, including asynchronous function calling | Supported; model documentation lists asynchronous function calling |
| Visual input | Supported | Supported |
| Live API | Supported | Supported |
| Thinking-level control | thinkingLevel is not exposed for this model |
Supports low, medium and high thinking levels |
The distinction is therefore more precise than “fast model versus smart model.” Both are capable live models. The standard version prioritizes responsive interaction, while Extended Thinking gives the system more room for background reasoning during complex tasks. [4] [5] [6]
Key Gemini 3.8 Live features
1. Low-latency, native voice interaction
Gemini 3.8 Live is optimized for real-time audio input and audio output. This makes it suitable for conversational assistants, customer-service agents, live copilots and other systems where pauses can make the experience feel unnatural. [4]
2. Background tool and API calls
Google highlights asynchronous function calling as a core feature. A voice agent can start an external task and continue the conversation instead of remaining silent until that task finishes. [1]
In practice, that can help with workflows such as checking an order, retrieving account information, updating a reservation or calling an enterprise service while the user continues speaking.
3. Visual grounding during live conversations
Gemini 3.8 Live can process visual inputs in near real time. Depending on the application, a spoken conversation can therefore be grounded in an image, video stream or camera context rather than audio alone. [1]
4. Automatic switching across 97 supported languages
Google says Gemini 3.8 Live can automatically detect and transition between 97 supported languages during a conversation. That is especially useful for multilingual users who naturally change languages mid-dialogue. [1]
5. Improved handling of alphanumeric information
Google also emphasizes stronger handling of identifiers such as order numbers, confirmation codes and similar alphanumeric information. This matters in production voice agents because one incorrect character can break a downstream workflow. [1]
6. Background reasoning with Extended Thinking
Gemini 3.8 Live Extended Thinking can process background reasoning and asynchronous tool calls while continuing to stream audio responses. This is intended for tasks that require more planning or multi-step problem solving without abandoning the live conversation. [5]
Where Gemini 3.5 Transcribe fits in
Gemini 3.5 Transcribe is separate from Gemini 3.8 Live. It is designed primarily for speech-to-text rather than two-way audio conversation. Google first released the Gemini 3.5 Transcribe models on August 26, 2026 and highlighted them again with its September voice-AI developer lineup. [2] [3]
Important API distinction: Google exposes two dedicated transcription
model IDs. gemini-3.5-transcribe is the non-streaming speech-to-text
model, while gemini-3.5-transcribe-live is the low-latency streaming
version delivered through the Live API.
[2]
Gemini 3.5 Transcribe capabilities
- Automatic language detection and transcription across 85+ languages.
- Multilingual code-switching without requiring manual language configuration.
- Custom vocabulary biasing for specialized terminology, names and uncommon words.
- Speaker diarization and word-level timestamps for supported pre-recorded workflows.
- Smart transcription features such as filler-word cleanup and formatting.
- A separate Live variant for low-latency streaming transcription.
Google cites Artificial Analysis measurements of approximately 4.0% average word error rate for streaming and 2.6% for non-streaming transcription. These are benchmark measurements, not guaranteed accuracy levels for every language, accent, microphone, environment or specialist vocabulary. [3]
| Model | Choose it when... | Primary output |
|---|---|---|
| Gemini 3.8 Live | You need a responsive, two-way conversational voice agent | Text and audio |
| Gemini 3.8 Live Extended Thinking | You need deeper reasoning during a live multi-step voice workflow | Text and audio |
| Gemini 3.5 Transcribe | You need accurate non-streaming speech-to-text | Text |
| Gemini 3.5 Transcribe Live | You need low-latency streaming speech-to-text | Text |
Gemini 3.8 Live pricing and API access
Google’s September 15, 2026 release notes list Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking as generally available through the Live API. Developers can try the models through the Gemini API and Google AI Studio. [2]
As of this article’s September 19, 2026 fact-check, Google lists standard paid-tier Gemini 3.8 Live-family pricing as follows: [7]
| Data type | Paid input | Paid output |
|---|---|---|
| Text | $0.75 per 1M tokens | $4.50 per 1M tokens |
| Audio | $3.00 per 1M tokens or about $0.005/min | $12.00 per 1M tokens or about $0.018/min |
| Image / video input | $1.00 per 1M tokens or about $0.002/min | Not listed as an image/video output mode |
Pricing can change. Always check Google’s current Gemini API pricing page before estimating production costs. Optional features such as search grounding or transcription can introduce additional charges under Google’s current billing rules.
Developer note: token limits
Google currently lists both Gemini 3.8 Live models with a 131,072-token input limit and 65,536-token output limit. [4] [5]
Who should use Gemini 3.8 Live?
Gemini 3.8 Live makes the most sense when the product needs an ongoing spoken conversation rather than a single voice command or simple transcription.
- Customer-service agents: retrieve information or trigger tools during a live call.
- Education and tutoring apps: discuss spoken questions together with visual material.
- Accessibility tools: combine audio interaction with visual context.
- Enterprise assistants: call internal APIs while maintaining a conversation.
- Travel and booking assistants: coordinate several steps or services during one voice session.
- Technical-support systems: reason over a user’s spoken explanation together with images or video.
When Gemini 3.5 Transcribe is the better choice
If your main requirement is simply to convert speech into text—for example, meeting transcription, captions, interview transcription or call analytics—the Transcribe models are a more direct architectural fit. A full audio-to-audio agent can add unnecessary complexity when conversational reasoning is not required.
Important limitations and cautions
- Benchmarks do not guarantee your application’s performance. Google reports strong evaluation results, but real-world quality depends on the task, language, audio conditions, tool design and system architecture.
- End-to-end latency is larger than model latency alone. Network conditions and external API calls can still introduce delays.
- Language support does not imply identical quality in every language. Developers should test the languages, accents and acoustic environments relevant to their actual users.
- Tool-using agents need safeguards. High-impact or irreversible actions should use confirmation, authorization and appropriate human oversight.
- Voice and camera data can be sensitive. Applications should minimize collection, request clear permissions and apply suitable retention and access controls.
- Some Gemini API features are not supported by these Live models. Google’s model pages currently list features such as code execution, file search and structured outputs as unsupported for Gemini 3.8 Live and Extended Thinking. [4] [5]
Why Gemini 3.8 Live matters
The bigger development is not simply a new Gemini version number. Gemini 3.8 Live represents the continuing shift from voice assistants that answer isolated commands to multimodal agents that can maintain dialogue while reasoning and performing external work.
That can change how software is designed. Instead of making users repeatedly navigate menus and forms, a voice-first application can potentially maintain context, understand visual information and coordinate tool calls while the conversation continues.
Bottom line
Gemini 3.8 Live is Google’s production-ready choice for most low-latency voice-agent experiences. Gemini 3.8 Live Extended Thinking adds more background reasoning for complex, multi-step interactions. For speech-to-text workloads, Gemini 3.5 Transcribe and Gemini 3.5 Transcribe Live provide dedicated non-streaming and streaming options respectively.
For developers, the practical question is therefore not “Which Gemini model is most powerful?” It is which model matches the workload: real-time dialogue, deeper voice reasoning, streaming transcription or non-streaming transcription.
Frequently asked questions
What is Gemini 3.8 Live?
Gemini 3.8 Live is Google’s generally available, low-latency audio-to-audio model for real-time voice agents and live dialogue. It supports multimodal input, interleaved reasoning, function calling and streamed audio output.
When was Gemini 3.8 Live released?
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026. Google’s public announcement was updated on September 17, 2026.
What is Gemini 3.8 Live Extended Thinking?
It is Google’s higher-reasoning audio-to-audio model for complex, multi-step problem solving during real-time voice interactions. It can perform background reasoning and asynchronous tool calls while streaming audio responses.
Does Gemini 3.8 Live support images and video?
Yes. Google’s model documentation lists text, images, audio and video as supported input types, with text and audio as output types.
How many languages does Gemini 3.8 Live support?
Google says Gemini 3.8 Live can automatically detect and transition between 97 supported languages during a conversation.
Is Gemini 3.5 Transcribe the same as Gemini 3.8 Live?
No. Gemini 3.8 Live is designed for two-way live voice interaction. Gemini 3.5 Transcribe is designed for speech-to-text. Google also provides a separate Gemini 3.5 Transcribe Live model for streaming transcription.
How many languages does Gemini 3.5 Transcribe support?
Google says the Gemini 3.5 Transcribe family automatically detects and transcribes more than 85 languages.
Is Gemini 3.8 Live generally available?
Yes. Google’s Gemini API release notes list both Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking as generally available from September 15, 2026.
Official sources
- Google — Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
- Google AI for Developers — Gemini API release notes
- Google — Intelligent transcription with Gemini 3.5 Transcribe
- Google AI for Developers — Gemini 3.8 Live model documentation
- Google AI for Developers — Gemini 3.8 Live Extended Thinking model documentation
- Google AI for Developers — Live API capabilities
- Google AI for Developers — Gemini Developer API pricing
Information and pricing were checked against the sources above on September 19, 2026. Product capabilities, model availability and pricing can change after publication.