Gemini 3.8 Live and 3.8 Live Extended Thinking — Google's most advanced live dialogue models

TL;DR

Google AI Studio announces two new speech-to-speech dialogue models in the Gemini 3.8 family: Gemini 3.8 Live (cost-efficient conversational agent with fluid dialogue and visual grounding) and Gemini 3.8 Live Extended Thinking (higher-intelligence variant that reasons and talks simultaneously for multi-step tasks). Extended Thinking claims #1 on Artificial Analysis’ Speech-to-Speech Quality Index at 82.6, plus 68.6% on τ-Voice, 35.1% on Sierra’s τ-Voice-banking, and 97.7% on Big Bench Audio. Both ship in the Gemini Live API today, with visual input, 97-language mid-conversation switching, background tool execution, and SynthID watermarking; no weights, no report, no architecture details.

Key claims

Gemini 3.8 Live Extended Thinking tops Artificial Analysis’ Speech-to-Speech Quality Index with a score of 82.6, and leads on agentic voice benchmarks with 68.6% on τ-Voice and 35.1% on Sierra’s τ-Voice-banking [tweet body].
Extended Thinking scores 97.7% on Big Bench Audio while remaining “highly competitive” on price against other frontier models [tweet body].
Gemini 3.8 Live places second in the Speech Agent Arena and is positioned as the cost-efficient sibling built for scale [tweet body].
Both models expand the Pareto frontier on ServiceNow EVA-Bench, balancing task accuracy with conversational quality [tweet body].
Gemini 3.8 Live processes visual input in near real-time and automatically detects/switches between 97 supported languages mid-conversation [tweet body].
Tool calls execute in the background while dialogue continues — the model can acknowledge a request and keep chatting while the task finishes [tweet body].
Extended Thinking reasons and speaks in parallel, using verbal cues like “Let me check that…” and live progress narration during multi-step background work [tweet body].
All generated audio carries an imperceptible SynthID watermark for downstream provenance detection [tweet body].

Method

Not disclosed. The announcement gives no architecture, parameter count, audio tokenizer/codec, training recipe, or latency figure with hardware. Both models are closed and API-only, delivered via the Gemini Live API in Google AI Studio, private preview in Gemini Enterprise, and rolled into Gemini Live, Search Live, Docs Live, Gmail Live, and Keep Live. The naming places them in the Gemini 3.8 family — the closed-frontier successor lineage to the Gemini 3.1 Flash cohort tracked in Gemini 3.1 Flash TTS: Google text-to-speech with scene direction and audio tags (announcement), Gemini 3.1 Flash Live: realtime voice and vision agent model (Google product announcement), and Nano Banana 2 / Gemini 3.1 Flash Image (Google product announcement). “Extended Thinking” implies a reasoning-heavy variant that runs internal deliberation concurrently with the speech-generation stream, but the mechanism (interleaved token streams? separate thinking head? scratchpad?) is not specified. The claim of background tool execution during ongoing dialogue implies asynchronous function-calling with a separate audio-generation loop, but again no implementation detail is given.

Results

Reported benchmark numbers for Gemini 3.8 Live Extended Thinking:

Artificial Analysis Speech-to-Speech Quality Index: 82.6 (rank #1 overall) [tweet body].
τ-Voice agentic task completion: 68.6% [tweet body].
Sierra τ-Voice-banking: 35.1% [tweet body].
Big Bench Audio: 97.7% [tweet body].
ServiceNow EVA-Bench: pushes the Pareto frontier for complex workflows (no absolute number in tweet) [tweet body].

For Gemini 3.8 Live: second place in the Speech Agent Arena; no absolute numbers, but positioned on the cost-efficient side of the frontier [tweet body]. No head-to-head against GPT-4o Realtime, OpenAI voice mode, or open speech-to-speech models. No latency-with-hardware figure. No parameter count.

Why it’s interesting

This is the speech-to-speech agent extension of the Gemini 3.8 family — the successor closed-frontier lineage after Gemini 3.7 Flash (Gemini 3.7 Flash Model Card) and the 3.1 Flash family whose voice component was Gemini 3.1 Flash Live: realtime voice and vision agent model (Google product announcement) (Gemini 3.1 Flash Live) and whose TTS component was Gemini 3.1 Flash TTS: Google text-to-speech with scene direction and audio tags (announcement). The “Extended Thinking” variant is the interesting design move: it packages Inference-Time Scaling into a live-audio format, with the model narrating its own progress during multi-step tool use rather than blocking on silent computation. That’s a direct challenge to the open-weight streaming-TTS-plus-LLM pipelines catalogued in Qwen3-TTS Family is Now Open Sourced: Voice Design, Clone, and Generation, Voxtral TTS — Mistral AI frontier open-weight text-to-speech model (announcement) (Voxtral TTS), and CosyVoice 2: Scalable Streaming Speech Synthesis with Large Language Models — none of which currently expose an integrated reasoning-during-speech surface.

The τ-Voice / Sierra τ-Voice-banking numbers are the most concrete data here: they’re voice-agent versions of Sierra’s τ-bench for tool-use, so 68.6% and 35.1% are the first published frontier benchmarks for voice-driven Tool-Use Agents rather than text-driven ones. That reframes speech models from “TTS quality contest” to “agent task completion under speech-only I/O” — the same shift text-mode agents went through around SWE-bench. Once open speech-agent stacks post τ-Voice numbers, this announcement is the closed-frontier baseline to point at.