Gemini 3.8 Live

Overview

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are Google's stable real-time audio models released on September 15, 2026. Both accept text, image, audio, and video inputs. Live sessions use audio as the response modality; enable output audio transcription when a text transcript is required. They are based on Gemini 3 Pro and support a 128K input context window with up to 64K output tokens.

The two endpoints solve different voice-agent problems. gemini-3.8-live prioritizes low-latency dialogue, direct commands, and simpler session handling. gemini-3.8-live-extended-thinking adds configurable background reasoning, parallel asynchronous tools, and spoken progress updates for complex work. The upgrade includes enough configuration and runtime-behavior changes that teams should treat it as a migration from Gemini 3.1 Flash Live Preview, not a model-ID-only switch.

What's New

Stable endpoints for production voice agents

Google now lists both gemini-3.8-live and gemini-3.8-live-extended-thinking as stable model codes. The standard model is the default choice for most low-latency voice agents, while Extended Thinking is the higher-reasoning option for live interactions that require planning or long-running tools.

Both models support text, image, audio, and video inputs with audio or text output. Their official model pages list 131,072 input tokens and 65,536 output tokens. They support Live API sessions, audio generation, search grounding, and function calling, but not caching, code execution, file search, image generation, structured outputs, URL context, or Batch API jobs.

More natural real-time interaction

Gemini 3.8 Live combines native audio output with visual context, interleaved reasoning, and background tool execution. Google says it can automatically detect and switch between 97 supported languages during a conversation. It can acknowledge a request and continue speaking while an API call runs, which is useful for customer support, device control, onboarding, and other AI agent workflows where silence makes the experience feel stalled.

Background reasoning with Extended Thinking

Extended Thinking reasons and speaks at the same time. It can narrate progress with brief conversational updates while it plans, calls multiple tools, and waits for asynchronous results. That makes it better suited to multi-step diagnostics, travel planning, research, code tutoring, and other voice tasks where tool calls may take several seconds.

This behavior changes the session contract. turnComplete: true ends the current utterance but does not prove that the overall task is finished. Clients must continue listening until interaction_status becomes IDLE, and every function declaration must use behavior: NON_BLOCKING.

Stronger published voice-agent results

Google reports that Gemini 3.8 Live Extended Thinking ranked first on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6. The launch post also reports 68.6% on tau-Voice, 35.1% on Sierra's tau-Voice-banking benchmark, and 97.7% on Big Bench Audio. Gemini 3.8 Live placed second in the Speech Agent Arena.

These are vendor-published or third-party benchmark results selected by Google. They are useful for shortlisting, but production teams should still test real accents, interruptions, tool latency, noisy audio, and completion accuracy in their own voice workflows.

Gemini 3.8 Live vs Extended Thinking

Start with Gemini 3.8 Live when fast turn-taking is the primary requirement. Choose Extended Thinking only when the extra reasoning and tool orchestration justify more client state, potentially more billed output tokens, and potentially longer task completion time.

Compared With Gemini 3.1 Flash Live Preview

Gemini 3.8 Live is Google's recommended replacement for gemini-3.1-flash-live-preview. It moves the voice stack from a legacy preview endpoint to a stable endpoint while expanding asynchronous workflows and changing several setup assumptions.

The standard 3.8 Live endpoint keeps the familiar rule that turnComplete: true closes the turn. Moving directly to Extended Thinking requires additional lifecycle work beyond this baseline migration.

Migration Guide

For a standard 3.8 Live migration:

1. Change the model string to gemini-3.8-live.
2. Remove thinking_level, thinking_config, and enable_affective_dialog from session setup.
3. Review function declarations because asynchronous NON_BLOCKING execution is now the default.
4. Treat turn_complete=true on client content as an unconditional interruption of active generation.
5. Remove attempts to disable proactive audio and limit video frames to the context the task actually needs.
6. Enable output audio transcription only if the application needs a text transcript, because transcription creates additional output tokens.

For Extended Thinking, also change the endpoint to gemini-3.8-live-extended-thinking, declare every tool as NON_BLOCKING, configure thinking_level as low, medium, or high, and keep the UI in a working state until interaction_status reports IDLE. MINIMAL thinking and synchronous blocking tools are not supported.

Availability & Access

Both models are available to developers through the Gemini API and Google AI Studio. Google also lists Google Cloud and Vertex AI distribution in the model card. Gemini 3.8 Live is rolling out in Search Live, while enterprise access is in private preview in Gemini Enterprise and planned for Gemini Enterprise for Customer Experience.

Extended Thinking is also rolling out through Gemini Live and selected Google Workspace surfaces. Google says Google AI Pro and Ultra subscribers can access it in Docs, while all Google AI subscribers can access it in Gmail and Keep. Availability, quotas, and enterprise rollout timing can differ by product and region.

Pricing & Cost Controls

Google groups both 3.8 Live endpoints under the same Standard Gemini API pricing:

Extended Thinking does not have a separate per-token rate, but its background reasoning tokens are included in output billing. Compare total cost per completed task rather than only the listed token rate. Context compression, selective video frames, short tool results, and careful reasoning-level selection can help control long-session costs.

Known Limitations

• Neither endpoint supports context caching, code execution, file search, structured outputs, URL context, Google Maps grounding, or Batch API processing.
• Extended Thinking rejects synchronous blocking tools and function scheduling configurations.
• Standard Live rejects thinking_level; Extended Thinking supports low, medium, and high but not minimal.
• Proactive audio is permanently enabled, which affects listening behavior and input-token costs.
• The model card lists a January 2025 knowledge cutoff and notes possible hallucinations, occasional slowness, and timeouts.
• All generated audio is marked with an imperceptible SynthID watermark.

Who Should Upgrade / Who Should Wait

Upgrade to Gemini 3.8 Live if:

• Your production voice agent still uses the legacy gemini-3.1-flash-live-preview endpoint.
• Low-latency spoken dialogue, visual context, language switching, and tool calls need to work in one stable API session.
• Your client can test always-on proactive audio, the changed thinking configuration, and asynchronous function behavior before switching traffic.

Choose Extended Thinking if:

• A single request regularly requires several tools, parallel retrieval, or multi-step reasoning.
• Spoken progress updates are preferable to silence while external systems respond.
• Your application can track interaction_status and enforce non-blocking tool execution.

Wait or stay on the existing model if:

• You rely on Gemini 3.1's configurable thinking_level, cannot accommodate always-on proactive audio, or plan to use Extended Thinking while still requiring synchronous BLOCKING tools.
• Predictable short turns matter more than complex task completion and you have not benchmarked 3.8 Live latency or cost.
• Your product cannot yet handle the new session-state rules or regression-test interruptions, transcripts, and noisy audio.

FAQ

Sources

• Gemini 3.8 Live and Extended Thinking launch
• Gemini 3.8 Live model page
• Gemini 3.8 Live Extended Thinking model page
• Thinking in the Live API
• Gemini 3.8 Audio model card
• Gemini Developer API pricing