Google Launches Gemini 3.8 Live and Extended Thinking Voice Models, Tops Speech-to-Speech Benchmark

Google today announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, positioning them as the company's "most advanced live dialogue models yet." The models succeed Gemini 3.1 Flash Live, which launched in March 2026, and are rolling out today to Gemini Live along with new conversational features across Gmail, Docs, and Keep.

Benchmark claims

According to Google, Gemini 3.8 Live Extended Thinking holds the #1 overall position on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6. The company also claims leading agentic task performance with 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking benchmark. On reasoning, Google reports a 97.7% score on Big Bench Audio.

The base Gemini 3.8 Live model, described as built for "scale and cost efficiency," claims second place in the Speech Agent Arena. Pricing for either model has not yet been disclosed.

What's new

Gemini 3.8 Live Extended Thinking is designed for high-complexity tasks and "reasons and speaks simultaneously," according to Google. The model uses verbal filler cues — such as "Let me check that…" — to acknowledge prompts while background processing occurs, and narrates progress during multi-step tasks. Google says this is intended to preserve conversational flow rather than forcing users to wait silently during longer operations.

The model powers three newly launched conversational products: Gmail Live (conversational search within email), Docs Live (draft generation and editing), and Keep Live (note creation). It is also beginning rollout to the standard Gemini Live app today.

The base Gemini 3.8 Live model focuses on "conversational intelligence with fluid dialogue and visual grounding," per Google. It processes visual inputs in near real-time and can execute tool and API calls in the background while conversation continues uninterrupted. The model supports mid-conversation language switching across 97 languages. Gemini 3.8 Live will also power the Search Live experience within Google's AI Mode.

Google has not disclosed parameter counts, training cutoff date, or context window size for either model.

What this means

This release signals Google's push to make voice AI feel less like a request-response system and more like a continuous conversation partner — a design pattern OpenAI and others have also pursued with real-time voice models. The emphasis on "speaking while thinking" and background task execution addresses a known weakness in voice assistants: dead air during processing breaks conversational immersion.

The benchmark claims, particularly the top Speech to Speech Quality Index ranking, are self-reported by Google and have not been independently verified by Artificial Analysis or third parties as of publication. The second-place finish in Speech Agent Arena for the base model, rather than first, suggests Google is positioning Extended Thinking as its flagship voice product while the standard model competes on cost and scale.

Deployment across Gmail, Docs, and Keep signals Google's strategy of embedding voice-driven AI directly into its productivity suite rather than treating it as a standalone assistant feature — a move that could pressure Microsoft's Copilot voice integrations if user adoption follows.