Gemini 3.8 Live Explained: Pricing, Voice Agents, Extended Thinking and API Limits

Gemini 3.8 Live is Google’s generally available low-latency, audio-to-audio model for real-time voice agents and live dialogue. Released on September 15, 2026, it accepts text, images, audio and video, and returns text or native audio. Google positions the standard model for fast conversational experiences, while Gemini 3.8 Live Extended Thinking is designed for harder multi-step requests that need background reasoning and asynchronous tool calls while the conversation continues.

The practical shift is not simply “better voice.” The new Live stack lets an agent listen continuously, keep the conversation moving while tools run, update session context during an interaction, and—on the Extended Thinking model—reason in the background without forcing a long silent pause.

Primary keyword: Gemini 3.8 Live. Search intent: informational / developer evaluation.

Google announced two related models: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The company describes the standard model as the default choice for most low-latency voice-agent experiences, with conversational intelligence, visual grounding, interleaved reasoning and asynchronous function calling. The Extended Thinking variant is aimed at complex workflows where a voice agent must keep talking naturally while planning, calling tools or waiting for external systems.

Both models are based on Gemini 3 Pro, according to Google DeepMind’s model card. They are natively multimodal: a live session can consume audio alongside text, images or video. That matters for scenarios such as guided troubleshooting, onboarding, field support and visual customer service, where the assistant needs to interpret what a person says and what a camera sees at the same time.

Availability is broad but not identical across the two variants. Google’s launch materials say Gemini 3.8 Live is rolling out through the Gemini API, Google AI Studio and consumer experiences including Search Live, while enterprise distribution also extends into Google’s business platforms. The Extended Thinking model is the higher-reasoning option and is exposed where Google supports that live reasoning workflow.

The two models share the same real-time conversational foundation, but they optimize for different interaction patterns. The standard model is the better fit when responsiveness is the priority: voice menus, live assistants, conversational search, helpdesk triage and other flows where a fast response matters more than deeper planning.

Extended Thinking is for requests where the agent must solve a problem rather than merely answer. Google’s Live API documentation describes a pattern in which the model can issue spoken progress cues while it performs background reasoning or asynchronous tool calls. The server exposes interaction status so an application can tell whether the model is still processing a larger task or has returned to an idle state.

For developers moving from the earlier gemini-3.1-flash-live-preview, the migration is more than a model-string swap. Google’s current model documentation highlights several behavioral changes that affect application architecture.

• Asynchronous function calling is the default. Tools can run without freezing the conversation, although blocking behavior remains available for compatibility.
• Client content can be updated across the session lifecycle. Applications can send additional user or model context while the session is active.
• Proactive audio is permanently enabled. Developers cannot turn it off on Gemini 3.8 Live, which has implications for listening behavior and cost.
• Video frames are included in turn coverage by default. Google recommends sending frames only when needed to control context size and cost.
• Native audio is the response modality. If an application needs a readable transcript, output audio transcription must be enabled.

Under the hood, the Live API is a persistent bidirectional session over WebSockets. Google’s GenAI SDK abstracts most of that transport, but applications still need to handle connection resets, session resumption, audio formats and cost-aware context management.

Google lists both Gemini 3.8 Live variants in the same Live pricing section. As of September 18, 2026, the paid Standard tier lists text input at $0.75 per 1 million tokens, audio input at $3.00 per 1 million tokens or about $0.005 per minute, and image/video input at $1.00 per 1 million tokens or about $0.002 per minute. Output is listed at $4.50 per 1 million text tokens and $12.00 per 1 million audio tokens or about $0.018 per minute.

Two details are easy to miss. First, transcription can add text-token charges when input or output transcription is enabled. Second, Google says proactive audio is permanently enabled on Gemini 3.8, so input costs continue while the API is listening. Long-lived voice agents therefore need explicit context and session-cost controls rather than assuming “silence is free.”

The most compelling use cases are ones where latency, natural speech and tool access matter at the same time. A conventional chatbot can wait several seconds and still feel acceptable; a voice interaction becomes awkward much faster.

Customer service and contact centers

A voice agent can identify intent, speak naturally, call account or CRM tools and continue the conversation while data is retrieved. Extended Thinking is especially relevant when the request spans several systems, such as checking an order, confirming policy rules and proposing the next action.

Guided troubleshooting

Because the model can accept video or image input alongside audio, an application can let a user show a device, error light or setup screen while describing the problem. That creates a different interaction model from a text-only support bot.

Employee onboarding and live assistance

Google’s launch examples emphasize real-time guidance. In an enterprise context, the useful part is not a generic spoken FAQ; it is a voice layer connected to tools, internal workflows and visual context, with strict permissions around what the agent can actually do.

Hands-busy interfaces

Field technicians, drivers, warehouse workers and accessibility-focused experiences can benefit from a conversational UI that does not demand constant typing. The business case depends on reliability and integration quality more than novelty.

Google’s model page advertises a large 131,072-token input window and 65,536-token output limit, but real-time sessions have operational constraints that developers need to handle separately.

Google’s session-management documentation states that, without context-window compression, audio-only sessions are limited to 15 minutes and audio-video sessions to 2 minutes. A WebSocket connection itself is limited to roughly 10 minutes. Session resumption can carry one logical session across multiple connections, and resumption tokens remain valid for a limited period.

Context-window compression is therefore not just an optimization for long conversations; it is part of the architecture. Developers can configure compression thresholds and sliding windows so old context is summarized or discarded before token history grows without bound.

Google DeepMind’s model card explicitly notes that Gemini 3.8 Live can still show general foundation-model limitations, including hallucinations, and may occasionally experience slowness or timeouts. That is important because spoken interfaces can make an answer feel more confident and immediate than the same text on a screen.

For production systems, the biggest risks are often architectural: excessive tool permissions, weak confirmation steps, sensitive data entering transcripts, or an agent taking irreversible actions from an ambiguous spoken request. Voice authentication is also a separate problem; a model understanding a caller does not prove the caller is authorized.

Google says generated audio is watermarked with SynthID. That helps provenance, but it does not replace application-level controls such as audit logs, least-privilege tool access, explicit confirmation for high-impact actions and clear disclosure that the user is speaking with an AI system.

Digital Pulse Brief’s practical recommendation is to treat the model as one layer in a voice-agent system, not as the security boundary. The safer architecture keeps identity, authorization, business rules and irreversible actions outside the model itself.