Google releases Gemini 3.8 Live and Extended Thinking speech models to production

Google is pushing its speech-to-speech models into production with two new releases designed to close the gap between natural conversation and complex task execution. Gemini 3.8 Live and 3.8 Live Extended Thinking are available now through the Live API and Google AI Studio as hosted models only — no self-hosting option and no open weights.

The launch covers two distinct roles. Gemini 3.8 Live targets scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding. Gemini 3.8 Live Extended Thinking is built for more complex workloads, adding multi-step reasoning that occurs while the model speaks. Google positions both as an efficient alternative to cascaded pipelines that chain ASR, LLM and TTS — an architecture that adds latency and failure points.

According to Google's figures, Extended Thinking takes the top spot on Artificial Analysis's speech-to-speech quality index with a score of 82.6. It leads on agentic task completion with 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking benchmark, and reaches 97.7% on Big Bench Audio, a reasoning test for audio models. The standard Gemini 3.8 Live placed second in the Speech Agent Arena human-preference leaderboard. On ServiceNow's EVA-Bench, Google reports the models push the Pareto frontier for complex workflows, balancing task accuracy against conversational quality.

The Live API exposes five core capabilities in the new models: asynchronous function calling that lets the model run APIs and tools in the background while audio responses continue streaming to the user; visual context for real-time visual input processing; alphanumeric accuracy for decoding verification codes, claim numbers and technical data — a common failure point in voice systems; multilingual support with automatic detection and switching across 97 languages mid-conversation while preserving accent consistency; and incremental content updates that merge real-time audio with structured data. Extended Thinking adds configurable background reasoning: the model reasons and speaks simultaneously, uses early verbal cues such as "let me check that" to acknowledge the request, then narrates step-by-step progress while long-running tasks execute. Google's demos show conversion of sketches and voice feedback into working React components, and coordination of multi-step bookings.

Both models are priced at $0.005 per minute of audio input and $0.018 per minute of audio output, an estimate based on $3 per million input tokens and $12 per million output tokens (approximately 11 and 44 shekels respectively). Developers can also build through integration partners that handle real-time streaming infrastructure: Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents. Google is also collaborating with Salesforce, Genspark and Lumeris, which cite latency, flow and tool calling as advantages. Sample applications are available on GitHub.