Gemini 3.8 Live: what it means for your teams
Google released two live voice models, and this page explains their availability, claims, and team fit.
Gemini 3.8 LiveTL;DR
- 01Gemini 3.8 Live handles real-time voice conversations and can call tools while people speak.
- 02Extended Thinking handles multi-step tasks and describes progress while it works.
- 03Google reports strong benchmarks, but the results need testing on your calls.
- 04Developers have access now. Enterprise access remains in private preview.
What shipped
Google DeepMind announced two live dialogue models on 15 September 2026. Gemini 3.8 Live focuses on scale and cost. Extended Thinking handles harder tasks and describes its progress while speaking.[1]
The models accept visual input in near real time. They can switch between 97 supported languages during a conversation. They can also call tools and APIs without pausing the conversation.[1]
Google says SynthID watermarks all audio generated by its AI products.[1]
Who can use it today
Availability stated in the announcement
Developers
Both models are rolling out in the Gemini API and Google AI Studio.
Enterprises
Both models are in private preview in Gemini Enterprise and are coming to Gemini Enterprise for Customer Experience. Extended Thinking is also coming to Workspace business customers.
Everyone
Gemini 3.8 Live powers Search Live. Extended Thinking powers Gemini Live, and Docs, Gmail and Keep for Google AI subscribers.
Google names Salesforce, Genspark and Lumeris as partners. Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents already use the Live API. These statements do not report customer results.[1]
Which of your teams should care
- Customer supportPhone and voice agents that look things up and act while the caller is still talkingStrongThe models support live customer conversations and tool calls.
- EngineeringBuilding the voice agents above on the Live APIStrongEngineers can build voice agents with the Live API.
- SalesVoice updates to the CRM and spoken meeting preparationPartialSales teams must connect the models to their CRM.
- FinanceLittle that a text assistant does not already doLowThe release adds little to current finance text tasks.
- LegalLittle today; recorded voice adds a records questionLowRecorded voice creates an extra records requirement.
- OperationsHands-busy work: field staff, onboarding walkthroughs that use the cameraPilotThe models accept voice and visual input during work.
Keep your current voice agent unless this model improves its results. Ask the vendor which model it uses and whether the change affects your price.
What to do this week
How to test it
- 1Choose one clear call type, such as an order status request. Count completed calls instead of minutes.
- 2Measure the share of calls that end without a human. Compare the result with Google's benchmark.
- 3Test interruptions, accents and a caller who changes language halfway through, because those are the claims.
- 4Decide what is recorded, where it is stored and who is told, before the first real customer call.
- 5Ask for enterprise pricing in writing. The announcement gives none.
The vendor's numbers
Google reports that Gemini 3.8 Live placed second in the Speech Agent Arena. Google also reports results for both models on ServiceNow's EVA-Bench. The footnote says Google ran that test on the Live API through the Gemini Enterprise Agent Platform.[1]
The announcement gives no price. It only calls the price competitive, so teams cannot budget from this announcement.[1]