Google Gemini 3.8 Live: Real-Time Voice AI for Enterprise
Google launched two new native audio AI models on September 15, 2026: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both are speech-to-speech models, taking audio in and returning audio out in near real time, designed specifically for voice agents that need to reason and execute tools without pausing the conversation. Gemini 3.8 Live Extended Thinking ranked first on the Artificial Analysis Speech-to-Speech Quality Index at launch with an overall score of 82.6, ahead of every other production voice AI model.
For enterprise teams building voice agents for customer experience, operations, or internal tools, this is the most significant Google voice AI release since the introduction of the Gemini Live API.
What Google Launched
The two models serve different positions in the cost-versus-capability spectrum.
Gemini 3.8 Live is built for scale and cost efficiency. It combines fluid dialogue with real-time visual grounding, automatic language detection across 97 languages, and background API calls while the audio stream continues. Google positioned it as the production workhorse for high-volume voice deployments. In independent Speech Agent Arena testing, it scored second overall.
Gemini 3.8 Live Extended Thinking is the high-complexity option. It adds configurable background reasoning: when a user makes a complex request, the model acknowledges it conversationally (“Let me check that…”), reasons through a multi-step plan while continuing to speak, and narrates its progress as each step resolves. This is not standard turn-taking. The model produces continuous audio output while inference is running, which is meaningfully different from the pause-and-respond pattern most voice agents use today.
Benchmark Performance
| Model | Artificial Analysis SSQ Index | Agentic (τ-Voice) | Banking (Sierra τ-Voice) | Big Bench Audio |
|---|---|---|---|---|
| Gemini 3.8 Live Extended Thinking | 82.6 (rank 1) | 68.6% | 35.1% | 97.7% |
| Gemini 3.8 Live | Second (Speech Agent Arena) | n/a | n/a | n/a |
| Previous Gemini Live models | Not ranked | n/a | n/a | n/a |
Sources: Google blog, Google DeepMind model card.
The τ-Voice benchmark evaluates agents completing real voice tasks, not just transcription accuracy. A 68.6% agentic task completion rate is the highest any production speech model has posted on that benchmark. The Sierra τ-Voice-banking score tests voice agents specifically in regulated financial service contexts, which explains why Gemini 3.8 Live has already attracted enterprise partners in banking.
Key Technical Capabilities
Background reasoning: The Extended Thinking model runs its reasoning pass in parallel with audio generation. It does not hold silence while thinking. Enterprise deployments that previously had to choose between intelligent responses and low-latency user experience no longer face that trade-off for most task types.
Async tool calls: Both models execute function calls and API integrations in the background while maintaining the audio stream. A voice agent can check a CRM, retrieve a document, or trigger a workflow while speaking, then weave the results back into the conversation. For enterprise use cases like customer support, logistics coordination, and internal operations, this collapses the round-trip cost of agentic voice interactions.
97-language switching: The model detects and transitions between 97 languages mid-conversation without a mode switch or a session restart. For global enterprise deployments with multilingual workforces or customer bases, this removes a major operational constraint.
Context window: Gemini 3.8 Live Extended Thinking supports 131,072 input tokens and 65,536 output tokens, matching the range needed for long enterprise calls with rich context, per the Google AI developer documentation.
Enterprise Deployment Path
Google is rolling both models into Gemini Enterprise and Gemini Enterprise for Customer Experience via private preview starting September 15. The path to general availability has not been given a specific date.
For Google Workspace customers, Extended Thinking is available immediately in Docs Live, Gmail Live, and Keep Live for Pro and Ultra subscribers. All Workspace subscribers can access it in Gmail and Keep.
Pricing via the Gemini API is $0.005 per minute for audio input and $0.018 per minute for audio output. That is a per-minute rate, not a per-token rate, which simplifies cost modeling for voice operations teams. A one-hour enterprise support call costs roughly $1.38 in model compute at those rates.
Early enterprise partners confirmed in the launch announcement include Salesforce, Genspark, and Lumeris, each of which highlighted latency, fluidity, and tool-calling capabilities as the primary reasons for adoption.
Developer Infrastructure
For teams building their own voice agents, both models are available immediately in the Gemini API and Google AI Studio via the Live API. Google’s launch also included integrations with six developer platform partners: Agora, Fishjam, LiveKit, LangChain, Pipecat, Vercel, and Vision Agents. These platforms manage real-time media streaming infrastructure, which is a substantial operational burden to build in-house, particularly for voice agent teams moving from a prototype to production.
The model code for Extended Thinking is gemini-3.8-live-extended-thinking on a stable release alias. Function calling is supported in async mode; structured outputs and batched API calls are not supported at launch.
Why This Matters for Enterprise AI Builders
Google’s previous voice work (Gemini 3.5 Transcribe) addressed the speech-to-text layer. The 3.8 Live family goes further: it eliminates the cascaded architecture pattern (speech-to-text, then LLM, then text-to-speech) that has been the standard for production voice agents. A cascaded system typically adds 800ms to 2 seconds of latency per turn and multiplies operational costs across three separate model calls.
The shift matters because enterprise voice AI has been one of the most contested spaces in 2026. OpenAI launched GPT-Live in July with strong early adoption; Anthropic added voice mode to Claude Opus. Google’s response is a native audio architecture rather than a voice wrapper around a text model, and the benchmark gap at launch is wider than most analysts expected.
For enterprise teams evaluating voice AI for customer experience, internal automation, or agentic workflows, the practical question is now whether the Extended Thinking model’s background reasoning capability justifies the premium over the base Live model for their specific task profile. Google’s benchmarks suggest the Extended Thinking model earns its position for complex multi-step tasks; the base Live model leads on latency and cost for simpler interactions.
Google’s enterprise rollout gates the Extended Thinking model to pilot customers in Gemini Enterprise initially. Teams wanting access before general availability should apply through the Gemini Enterprise private preview program.
If your team is building or deploying AI-native voice agents, book a call with Enera to map the build options against your stack and compliance requirements.