Skip to content
NewsModel release

Google Gemini 3.8 Live Extended Thinking tops Speech-to-Speech Quality Index with score of 82.6

· by Pondero Newsdesk

The short version

Google released Gemini 3.8 Live and 3.8 Live Extended Thinking on September 15, 2026. The Extended Thinking variant scored 82.6 on Artificial Analysis' Speech-to-Speech Quality Index and 68.6% on the tau-Voice agentic task completion benchmark, the highest marks on both leaderboards at launch.

Google Gemini 3.8 Live Extended Thinking tops Speech-to-Speech Quality Index with score of 82.6

Gemini 3.8 Live Extended Thinking scored 82.6 on Artificial Analysis' Speech-to-Speech Quality Index, per Google's September 15 announcement, the highest mark recorded on that leaderboard at the time of launch. The model adds inline reasoning to audio-to-audio voice interactions, letting it work through multi-step problems before speaking rather than requiring a separate text LLM chained into the pipeline.

What

Google launched two real-time voice models on September 15, 2026: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both are audio-to-audio systems, meaning speech goes in and speech comes out without a discrete text-processing step exposed to the developer.

The base Gemini 3.8 Live processes visual inputs in near real-time and supports 97 languages with mid-conversation switching. It ranked second in the Speech Agent Arena for cost efficiency, per Google's blog post. The Extended Thinking variant adds a reasoning pass before the model responds aloud. During that pass, it emits verbal acknowledgments such as "Let me check that..." rather than falling silent, and it can call background tools and APIs without interrupting the dialogue. Per Google, Extended Thinking scored 68.6% on the tau-Voice agentic task completion benchmark, 35.1% on Sierra's tau-Voice-banking benchmark, and 97.7% on Big Bench Audio reasoning.

All AI-generated audio from both models carries SynthID watermarking, a technique Google uses to tag output without audible alteration.

Availability on launch day: both models via the Gemini API and Google AI Studio for developers. Extended Thinking is in private preview through Gemini Enterprise; consumer access arrived through the Gemini app, Search Live, and Google Workspace (Docs, Gmail, Keep). Google did not announce pricing for the new Live API tier.

Why it matters

Voice agent deployments have typically required three separate components: a speech-to-text transcription model, a reasoning LLM, and a text-to-speech synthesizer. Each hop adds latency and a failure point. A native audio-to-audio model that retains reasoning capability and can invoke tools mid-conversation addresses that architecture directly.

The Artificial Analysis Speech-to-Speech Quality Index is a third-party benchmark, not a Google-administered test. An 82.6 score at the top of that leaderboard is a verifiable external signal. The 68.6% tau-Voice agentic task completion score carries similar weight: tau-Voice tests whether a model can complete realistic voice-agent tasks end-to-end, not just transcribe correctly.

The practical constraint for most developers right now is the availability gap. The Extended Thinking variant that holds those benchmark marks sits behind a private preview gate on Gemini Enterprise. Developers with standard Gemini API access can use the base Gemini 3.8 Live today, which Google positions for cost efficiency rather than maximum task complexity.

What to watch next

Two milestones are worth following. First, whether Google opens Extended Thinking to standard Gemini API tiers and what rate it carries. Second, whether the 82.6 Speech-to-Speech Quality Index lead holds once competing voice models receive updates, particularly given the pace of releases from OpenAI's real-time API team.

Sources