Power your agents: Gemini 3.8 Live with Live Avatar is now generally

Google shipped Gemini 3.8 Live with Live Avatar into general availability on September 25, 2026, moving the multimodal streaming model out of preview and opening it to production workloads. The release follows Google's announcement of Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking last week on Google Cloud Blog, and marks the point where developers can build and ship products on the model without a preview sign-up or early-access waitlist.

This article covers what changed between the preview and the GA release, how to wire Gemini 3.8 Live and Live Avatar into an agent stack, and what builders need to verify before going live. If you are maintaining a voice or video agent on an older Gemini model, or evaluating real-time multimodal APIs for the first time, the sections below give you the concrete steps.

What Power your agents shipped

Google marked Gemini 3.8 Live and its Live Avatar capability as generally available, meaning both the streaming audio model and the rendered avatar surface are now production-grade with associated SLAs. Gemini 3.8 Live Extended Thinking, which was announced alongside the base model last week, allows the model to reason through multi-step queries before returning a spoken response. The Live Avatar layer maps the model's audio output to a photorealistic or stylized on-screen persona in real time, giving voice agents a visual presence without a separate rendering pipeline.

The GA designation covers the core Live API endpoints in Google Cloud, the Extended Thinking variant, and the Live Avatar rendering service. Developers who joined the preview programme have their existing API keys honored at GA; new callers follow the standard Google Cloud credential flow through the project console.

What changed for builders in Power your agents

The source announcement describes the GA promotion itself but does not publish a comparison table of preview versus GA latency, pricing, or rate limits. Builders should pull current quota and pricing from the Google Cloud console under Vertex AI → Generative AI, where the Gemini 3.8 Live line items will now carry production-tier quotas rather than preview allocations.

The distinction that matters operationally is SLA coverage: preview endpoints carry no uptime commitment, while GA endpoints fall under the Google Cloud Service Level Agreement. If your agent has an SLA of its own to customers, this is the change that closes the gap.

How to install or upgrade Power your agents

Google distributes Gemini 3.8 Live access through the Vertex AI SDK and through the Google GenAI SDK. To update the Python SDK to the version that surfaces the GA model identifiers, run:

pip install --upgrade google-generativeai

For the Vertex AI path:

pip install --upgrade google-cloud-aiplatform

To point an existing agent at the GA model, set the model string in your configuration:

model = "gemini-3.8-live"

For Extended Thinking:

model = "gemini-3.8-live-extended-thinking"

Live Avatar is enabled by setting the avatar configuration on the session object as documented in the Gemini Live API reference; no separate package install is required. Confirm the active model by printing model.name from the client response before routing production traffic.

Gotchas and compatibility in Power your agents

The Live API uses a WebSocket session model rather than the standard request-response shape of the Gemini text endpoints. Agents built on the REST generateContent path will not drop in without a session management layer that holds the socket open, handles turn-taking, and reconnects on disconnection. Builders moving from preview to GA should audit their reconnect logic, because GA quotas and rate limits may differ from what the preview environment enforced.

Live Avatar renders output tied to the audio stream, so any latency introduced by intermediate processing — transcription buffers, function-call round trips, or tool latency inside Extended Thinking — delays the avatar animation proportionally. Measure end-to-end frame latency in a staging environment before enabling Live Avatar in production; the visual lag is more noticeable to end users than audio-only latency at the same millisecond delta.

What to watch after Power your agents

Google announced Extended Thinking alongside the base Live model last week, and the GA release covers both surfaces simultaneously. Watch the Google Cloud release notes for quota increases that typically follow a GA launch in the first 30 to 60 days, as preview caps are often conservative relative to what the production infrastructure supports.

Pricing for streaming audio tokens and for Live Avatar rendering minutes will be the cost driver for voice agents at scale. Neither figure appeared in the GA announcement text available at publication time; verify both in the Google Cloud Pricing calculator under Vertex AI before committing to a high-volume deployment. The Extended Thinking variant, which runs additional reasoning steps, will consume more tokens per turn than the base model, so instrument per-session token counts from day one.