Gemini 3.8 Live and Extended Thinking: Voice-Agent Models, Benchmarks and Availability
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026, expanding its real-time voice stack with visual grounding, background tool execution and automatic switching across 97 supported languages. Both models are rolling out to developers through the Gemini API and Google AI Studio.
The two models target different deployment profiles. Gemini 3.8 Live is optimized for scale and cost efficiency, while Gemini 3.8 Live Extended Thinking adds deeper multi-step reasoning for complex voice-agent workflows. Google says Extended Thinking can reason while speaking and narrate progress while asynchronous tools run in the background.
Google reports that Extended Thinking scored 82.6 on Artificial Analysis' Speech to Speech Quality Index, 68.6% on τ-Voice, 35.1% on Sierra's τ-Voice-banking benchmark and 97.7% on Big Bench Audio. These are launch benchmark results reported by Google; production evaluation should use the latency, task-completion and tool reliability requirements of the intended application.
Gemini 3.8 Live vs Extended Thinking
| Capability | Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking |
|---|---|---|
| Primary target | Scaled, cost-efficient live dialogue | Complex multi-step voice workflows |
| Voice interaction | Near-real-time | Near-real-time with deeper reasoning |
| Visual grounding | Yes | Yes |
| Background tools/API calls | Yes | Yes |
| Automatic language switching | 97 supported languages | 97 supported languages |
| Developer access | Gemini API, Google AI Studio | Gemini API, Google AI Studio |
| Enterprise rollout | Gemini Enterprise private preview | Gemini Enterprise private preview |
| Consumer rollout | Search Live | Gemini Live; selected Workspace subscriber experiences |
The architecture is particularly relevant to agents that need to maintain a conversation while external work continues. A support agent can acknowledge a request, call a CRM or ticketing API, keep the session active and incorporate the returned result without forcing the user through a sequence of disconnected voice turns.
Real-time visual grounding and 97-language switching
Gemini 3.8 Live can process visual input alongside a live conversation. Google demonstrates the model using visual context for employee onboarding and interacting with a chess position while maintaining spoken dialogue.
The model also detects and transitions between 97 supported languages during a conversation. That reduces application-side routing requirements for multilingual deployments where users can change languages within the same session.
Language count alone is an incomplete deployment metric. Contact-center and enterprise teams should evaluate recognition accuracy, pronunciation, latency, domain terminology and tool-call completion in the specific languages and accents they expect in production.
Background tools change the voice-agent interaction model
Traditional voice systems often expose the latency of each backend action: the user asks for something, the system waits for a tool response, and the conversation resumes afterward. Gemini 3.8 Live can execute tools and API calls in the background while continuing the conversation.
Extended Thinking builds on that behavior for longer workflows. Google describes the model as reasoning and speaking simultaneously, with progress narration while multi-step background tasks execute. Demonstrations include transforming sketches and spoken feedback into React components and coordinating multi-step bookings with asynchronous function calls.
For developers, this shifts several engineering requirements from raw model quality toward orchestration quality. A production implementation needs clear handling for tool timeouts, duplicate actions, cancellation, stale results and user corrections that arrive while a background operation is still running.
Benchmark results
Google's launch material reports the following results for Gemini 3.8 Live Extended Thinking:
| Benchmark | Reported result | Measures |
|---|---|---|
| Artificial Analysis Speech to Speech Quality Index | 82.6 | Overall speech-to-speech model quality |
| τ-Voice | 68.6% | Agentic voice task completion |
| Sierra τ-Voice-banking | 35.1% | Banking-oriented agent task completion |
| Big Bench Audio | 97.7% | Audio reasoning |
Google says Gemini 3.8 Live also placed second in the Speech Agent Arena, a preference-based evaluation. ServiceNow's EVA-Bench testing places the models on a Pareto frontier balancing task completion and conversational experience; Google notes that this evaluation ran through the Live API on the Gemini Enterprise Agent Platform.
Benchmark leadership can vary with benchmark revisions, model updates and serving configurations. The more durable engineering question is whether a model reaches the required task-completion rate at acceptable end-to-end latency and cost under the application's actual tool chain.
Where the models are available
Google began the rollout on September 15, 2026.
For developers, both models are available through the Gemini API and Google AI Studio. Gemini 3.8 Live is in private preview for Gemini Enterprise and is planned for Gemini Enterprise for Customer Experience. Extended Thinking is also in Gemini Enterprise private preview, with expansion planned for Customer Experience and Google Workspace business customers.
On Google's consumer surfaces, Gemini 3.8 Live is rolling out through Search Live. Extended Thinking is rolling out through Gemini Live and selected Workspace Live experiences, with access varying by Google AI subscription tier and product.
All audio generated by Google's AI products is watermarked with SynthID, according to the launch announcement.
What to test before deploying a voice agent
A useful evaluation should measure the complete session instead of treating speech quality as the only success metric.
Task completion: test representative workflows that require several tool calls, authentication boundaries and state changes.
End-to-end latency: measure time from the end of a user utterance to useful spoken acknowledgement, plus total time until the requested backend action completes.
Interruption and correction handling: users routinely change instructions while an agent is speaking or while a tool call is running. Test cancellation and revised-intent behavior explicitly.
Tool reliability: measure successful calls, malformed arguments, retries, duplicate actions and recovery after API failures.
Multilingual performance: validate the languages, accents and code-switching patterns used by the target population rather than extrapolating from the supported-language count.
Visual-session behavior: applications using camera or screen context should test how quickly changed visual information is incorporated and how sensitive task completion is to image quality.
Bottom line
Gemini 3.8 Live moves Google's voice-agent stack toward continuous multimodal sessions in which conversation, visual context and tool execution can proceed concurrently. Extended Thinking adds a higher-reasoning tier for workflows that require several dependent actions while preserving live dialogue.
The launch benchmarks are strong, especially the 82.6 Speech to Speech Quality Index and 68.6% τ-Voice result reported for Extended Thinking. Deployment decisions should pair those model-level results with application-level testing of latency, tool completion, interruption handling and multilingual accuracy.