Gemini 3.8 Live

// Overview

Gemini 3.8 Live is Google's audio-to-audio model for the Gemini Live API, released to general availability on 15 September 2026 as the model id `gemini-3.8-live`. Google's documentation calls it the default option for most low-latency voice agent experiences and real-time dialogue, and positions it as the variant to reach for when reasoning delays would be felt in the conversation.

It consumes a continuous stream rather than discrete requests: text, images, audio and video go in over a WebSocket, and text and audio come back out. Google's model card states that both 3.8 Live variants are based on Gemini 3 Pro, with a 128K-token input window (131,072 tokens) and a 64K output limit (65,536 tokens), and a knowledge cutoff of January 2025. The Live API caps a single audio-only session at 15 minutes and an audio-plus-video session at 2 minutes.

Speech behaviour is configurable rather than fixed. The model can use any voice available to Google's text-to-speech models, selected through `speechConfig`, and covers the 97 languages the Live API documents — switching between them naturally mid-conversation rather than being pinned to one at session start. Function calling is asynchronous, with `NON_BLOCKING` behaviour as the default plus function scheduling, so a tool call does not stall the audio stream.

Pricing is metered two ways on Google's pricing page: per million tokens, or per minute of media. Text is $0.75 in and $4.50 out per million tokens; audio is $3.00 in and $12.00 out per million tokens, equivalently $0.005 and $0.018 per minute; image and video input is $1.00 per million tokens, or $0.002 per minute. The Extended Thinking sibling is billed at the same rates.

// Pricing

Audio: $3.00 in / $12.00 out per 1M tokens, or $0.005 / $0.018 per minute. Image and video input: $1.00 per 1M tokens, or $0.002 per minute.

Pricing source ↗

// Strengths

• Answers without a reasoning pause, which is why Google names it the default for low-latency voice agents
• Native audio-to-audio: streaming audio, video, images and text in, text and audio out, over one WebSocket session
• 97 documented languages, with natural switching between them during a conversation rather than a fixed session language
• Any text-to-speech voice can be selected via `speechConfig`, so the same model covers different product personas
• Asynchronous (`NON_BLOCKING`) function calling with scheduling, so tool work does not stall the spoken reply

// Best for

• Reach for it for phone-style voice agents and support lines where a pause before the first word reads as a fault.
• Reach for it for live multimodal help — a user pointing a camera at something while talking — inside the 2-minute audio-plus-video session limit.
• Reach for it for multilingual assistants that need to follow a caller who switches language mid-sentence.
• Reach for it when a voice front end has to trigger lookups or actions mid-conversation via non-blocking function calls.

// How to access

// Gemini Live (real-time audio) — every version

The full lineage of the Gemini Live (real-time audio) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.