Google ships Gemini 3.8 Live for production voice agents

Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on 15 September 2026 as two speech to speech native models aimed at production voice agents. The distinction is deliberately operational. Gemini 3.8 Live is presented as the option for scale and economic efficiency, where a company wants many live conversations to run at predictable cost. Gemini 3.8 Live Extended Thinking is presented as the option for harder tasks that need multi step reasoning before the system answers, calls a tool, or changes direction. Both models are available through the Gemini Live API and Google AI Studio, which means teams can test them in hosted form immediately, but they are not open weights and they are not offered for self hosting.

Two hosted models for live voice agents

The release matters because voice agents fail in different ways from text chatbots. A text system can pause visibly while it retrieves data; a phone or meeting agent has to manage silence, interruptions, turn taking and tool results without making the conversation feel broken. Google says asynchronous function calling is meant to address part of that problem. The model can start tools and APIs in the background while the dialogue continues, instead of forcing the caller to wait for every external action to finish. That is not a full production guarantee, but it is a concrete design choice for support agents, booking flows, internal help desks and any voice workflow that must consult live systems during a call.

• Gemini 3.8 Live targets scale and cost efficiency for production voice agents.
• Gemini 3.8 Live Extended Thinking targets complex tasks and multi step reasoning.
• Both models are speech to speech native and were made available on 15 September 2026 through the Gemini Live API and Google AI Studio.
• Google says asynchronous function calling lets tools and APIs run in the background while conversation continues.
• The models are hosted services, not open weights, so customers cannot self host them.

What Google says the models can handle

Google also positions the models as multimodal live systems rather than voice only engines. They can process visual inputs in near real time, which matters for agents that need to look at a screen, a product, a document or a workspace while speaking. Google says the models can detect and automatically switch between 97 languages, a capability aimed at multilingual calls where users do not stay inside one language. The company also says it has improved precision for codes and alphanumeric data, a practical point for voice agents that must hear order numbers, account identifiers, booking references or short verification strings. Those are usually the moments where an impressive demo becomes expensive in production if the agent mishears one character.

Benchmarks are announced results, not a general validation

Google reports several benchmark results for Gemini 3.8 Live Extended Thinking: 82.6 on the Speech to Speech Quality Index, 68.6 percent on tau-Voice, 35.1 percent on Sierra tau-Voice-banking and 97.7 percent on Big Bench Audio. Google also says Gemini 3.8 Live ranks second on Speech Agent Arena. These figures are useful signals about what Google believes it has improved, but they should be read as announced results, not as a broad independent validation of every production setting. Voice agent quality depends on latency, acoustics, telephony, interruption handling, tool reliability and supervision, none of which is fully captured by a single score.

MarkTechPost reports pricing of 0.005 dollar per minute for audio input and 0.018 dollar per minute for audio output. The same report describes that estimate as based on 3 dollars per million input tokens and 12 dollars per million output tokens. For buyers, this matters less as a headline rate than as a starting point for a model of real usage. A production agent pays for greetings, failed turns, repeats, transfers, abandoned sessions and monitoring traffic. The lower cost profile promised for Gemini 3.8 Live will only show up if the whole conversation loop is measured, including any background tools launched through asynchronous function calling.

Infrastructure and deployment paths

Google names Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents as real time infrastructure partners around the release. The deployment path is also staged. The API and Google AI Studio are available immediately; private previews are planned in Gemini Enterprise; and availability in Search Live, Gemini Live and selected Workspace experiences depends on the relevant offerings. Google also says every sound generated by its AI products carries the imperceptible SynthID watermark. That does not remove the need for disclosure or audit logs in business settings, but it gives buyers one more control to include in their governance checklist.

Before moving a voice agent into production, buyers should test end to end latency, interruptions, recovery after errors, real cost per resolved task, traceability of tool calls and human supervision. That is analysis, not a claim announced by Google. For English speaking companies, the immediate change is that a hosted speech native Gemini line can now be evaluated as production infrastructure rather than as a demo layer. The right pilot is therefore not a scripted showcase. It is a measured call flow with noisy input, mixed accents, account numbers, tool failures, escalation rules and a clear decision about when a human must take over.