Google Launches Gemini 3.8 Live Models That Can Reason While They Talk

Google launched Gemini 3.8 Live Extended Thinking on Sept. 15, a voice model that continues reasoning through multistep tasks and runs tools in the background while audio streams, alongside the lower-latency Gemini 3.8 Live. Both models are generally available through the Gemini API and Google AI Studio, support 97 languages, and carry audio pricing of $3 per million input tokens (~$0.005/min) and $12 per million output tokens (~$0.018/min). Extended Thinking scores 82.6 on Artificial Analysis' Speech-to-Speech Quality Index and 68.6% on its agentic task-completion measure, and requires developers to track interaction_status instead of turnComplete and to use asynchronous, nonblocking function calls.

Voice assistants often fall silent when a request gets complicated. Google launched Gemini 3.8 Live Extended Thinking on Sept. 15 to keep the conversation going while the model reasons through multistep tasks and runs tools in the background. The model arrived alongside the lower-latency Gemini 3.8 Live. For developers, Extended Thinking changes a basic assumption of voice-agent design: the end of a spoken response no longer necessarily means the underlying task is finished. Both models are generally available through the Gemini API and Google AI Studio, according to Google’s Gemini API release notes https://ai.google.dev/gemini-api/docs/changelog?authuser=6 . They join Google’s expanding Gemini model family https://www.techrepublic.com/article/news-google-ai-models-explained/ , which spans general-purpose reasoning, multimodal applications and specialized AI workloads. | Feature | Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking | |---|---|---| | Primary use | Low-latency, real-time voice interactions | Complex, multistep voice tasks that require deeper reasoning | | Reasoning | Interleaved reasoning optimized for fast responses | Higher background reasoning during live audio interactions | | Tool use | Async function calling by default; blocking mode supported for backward compatibility | Async, nonblocking function calls only | | Best fit | Fast conversational agents and high-volume voice experiences | Agents that need deeper analysis, planning, or longer-running tools | | Availability | Generally available through the Gemini API and Google AI Studio | Generally available through the Gemini API and Google AI Studio | | Audio pricing | $3 per million input tokens ~$0.005/min ; $12 per million output tokens ~$0.018/min | Same standard audio rates | How background reasoning changes the voice loop Google’s Live API thinking guide https://ai.google.dev/gemini-api/docs/live-api/thinking?authuser=5 tells developers using Extended Thinking to track interaction status rather than relying on turnComplete . A turnComplete: true signal can arrive after an intermediate spoken update even while reasoning or a tool call continues. During that work, interaction status remains IN PROGRESS and switches to IDLE when the interaction is finished. Extended Thinking also requires asynchronous function calls, allowing tools to keep working while audio continues to stream; blocking calls are not supported. That architecture fits Google’s broader push toward agents that work across software and services. Gemini already has connections to apps, files and MCP servers https://www.techrepublic.com/article/news-gemini-spark-app-integrations/ , although those integrations come with their own access-control and governance considerations. The models accept text, images, audio and video as inputs and return text and audio. Google says Gemini 3.8 Live can automatically switch among 97 supported languages during a conversation, according to its Sept. 15 launch announcement https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/ . That could be particularly useful for multilingual customer-service and enterprise deployments across Asia-Pacific. Extended Thinking scores 82.6 on Artificial Analysis’ Speech-to-Speech Quality Index . The Artificial Analysis leaderboard https://artificialanalysis.ai/speech-to-speech also records a 68.6% score on its agentic task-completion measure. Google’s Gemini API pricing documentation https://ai.google.dev/gemini-api/docs/pricing?authuser=451499271 lists audio input at $3 per million tokens, or about $0.005 per minute, and audio output at $12 per million tokens, or about $0.018 per minute. Google has also introduced pay-as-you-go pricing and spending caps https://www.techrepublic.com/article/news-google-gemini-enterprise-pay-as-you-go-pricing/ for Gemini Enterprise as organizations weigh broader AI deployment costs. The production trade-offs behind smoother conversations The models are stable and generally available, but Google’s broader Live API remains in preview https://ai.google.dev/gemini-api/docs/live-api/capabilities?authuser=5 . Organizations moving beyond pilots will need to account for that status when evaluating integrations and support requirements. Reliability still requires testing. Google’s Gemini 3.8 Audio model card https://deepmind.google/models/model-cards/gemini-3-8-audio/ says both models can hallucinate and may occasionally experience slowness or timeouts, making retry, verification and failure-handling logic important for transactional applications. Voice output also cannot safely double as a completion signal. Interfaces and downstream systems should wait for the appropriate state before treating a booking, lookup or other action as finished, even when the model sounds as though it has already responded. Let us teach you How to Talk to AI for free Try our six-minute course at The Neuron Academy and learn a few simple ways to write better prompts and get more useful results from AI, or browse our other AI course for free for seven days. Check out all the lessons here →