Google brings extended reasoning to Gemini’s live voice models

Google has announced two voice models: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both can process streaming audio and visual input, maintain a spoken conversation, and call tools while a session continues.

The release gives developers two workload profiles. Gemini 3.8 Live targets high-volume applications where cost and responsiveness matter, while Extended Thinking adds multi-step reasoning and spoken progress updates for longer tasks. That division resembles the Flash and Pro tiers elsewhere in the Gemini lineup, adapted for real-time audio.

Two models, two workload profiles

Google maps the base model to customer support, live guidance, and other latency-sensitive interactions. Extended Thinking targets tasks such as debugging, multi-step bookings, and transforming a visual sketch into working code through spoken feedback.

Higher scores, limited disclosure

Google reports that Extended Thinking ranks first on Artificial Analysis’ Speech to Speech Quality Index and leads several tests of audio understanding and agentic task completion. The company also says the base model placed second in Speech Agent Arena.

ServiceNow’s EVA-Bench results assess whether voice agents can complete complex workflows while preserving conversational quality. Google’s charts place both models on the benchmark’s Pareto frontier, meaning each offers a competitive balance between those two measurements.

The published results do not include every configuration detail needed for independent comparison, including session settings, latency distributions, and tool definitions. Benchmark rankings can also change as evaluators add models and update tests.

The models coordinate speech, visual input, and tool execution within one live session. That architecture supports three capabilities with direct consequences for application design:

  • Near-real-time visual grounding: The models can use a camera feed or shared screen as conversational context, allowing a user to point at an object, interface, or error while speaking.
  • Automatic language switching: Google says the models can detect and switch among 97 supported languages during a conversation without a manual setting or restarted session.
  • Background tool calls: The models can invoke APIs while maintaining the dialogue, then incorporate returned data when the call completes.

Extended Thinking adds spoken status updates during multi-step work. A model might acknowledge that it is checking a booking, report that it is waiting for availability, and continue after the API responds. These updates summarize task status while the model keeps its hidden reasoning private.

Concurrent tool use reduces the silent gaps that often make voice interfaces appear disconnected. It also creates engineering obligations around cancellation, idempotency, stale results, and mid-call corrections. If a user changes a date while a booking request is running, the application still needs to cancel or reconcile the earlier call.

Workflows that fit the design

Google’s demonstrations focus on tasks that combine conversation with visual context or external systems. Suitable deployment targets include:

  • Customer-support agents that retrieve account records while speaking with a caller
  • Call-center systems that follow users who switch languages during a conversation
  • Field-service assistants that answer questions about a live camera feed
  • Programming tutors that inspect code, call development tools, and explain each action
  • Voice-driven pair programmers that turn sketches into React components
  • Booking agents that coordinate several asynchronous API calls
  • Onboarding assistants that combine screen context with account configuration tools

These applications still require product-specific controls. Developers need to define tool permissions, validate arguments, protect sensitive data, and ensure that spoken status messages match the actual state of each request.

Rollout spans APIs and Google products

Developers can access the models through the Gemini API and Google AI Studio. Consumer and enterprise availability varies by product and subscription:

Google presents migration for existing Live API clients as a model-name change. A minimal asynchronous Python session follows the same connection, send, and receive pattern described in the Live API docs:

from google import genai

client = genai.Client()
config = {"response_modalities": ["AUDIO"]}

async with client.aio.live.connect(
    model="gemini-3.8-live-extended-thinking",
    config=config,
) as session:
    await session.send(
        input="Walk me through debugging this stack trace",
        end_of_turn=True,
    )

    async for response in session.receive():
        if response.data:
            handle_audio(response.data)

The sample assumes that authentication, audio playback, error handling, and the handle_audio function already exist. Teams should confirm the current model identifier and SDK method signatures during the rollout, then test interruption handling, session limits, reconnect behavior, and tool-call concurrency before production use.

Agora, Fishjam, LiveKit, Pipecat, Vercel, and Vision Agents provide integrations for the Gemini Live API. Those platforms can manage media transport concerns such as WebRTC connections, jitter buffering, echo cancellation, and stream recovery.

Native audio changes the latency budget

Traditional voice agents often connect automatic speech recognition, a text language model, and text-to-speech synthesis. Every stage adds latency and can lose information about timing, tone, interruptions, or speaker intent. Native speech models process and generate audio within a single model session, reducing the number of handoffs.

OpenAI’s Realtime API, Advanced Voice Mode, and speech platforms built around ElevenLabs pursue related low-latency architectures. Google’s differentiator in this release is the combination of native audio, visual grounding, concurrent tool calls, and an extended-reasoning tier within the same API family.

Spoken progress updates can accommodate longer tool chains without leaving the user in silence. Product teams may need less filler-audio logic and speculative prefetching, though actual savings will depend on measured first-audio latency, tool duration, and the quality of interruption handling.

Deployment gaps remain

Google has not published several figures needed for production planning:

  • End-to-end latency percentiles and time to first audio
  • The latency difference between the base and Extended Thinking models
  • Live-session context limits and maximum session duration
  • Per-minute input and output pricing
  • Rate limits, regional availability, and production service-level commitments
  • Detailed behavior for interruptions, overlapping speakers, and failed tool calls

Enterprise access also remains restricted in parts of the product line, which limits immediate deployment through Customer Experience and Workspace channels. API access offers an earlier path, subject to the quotas and terms attached to each account.

Google says all audio generated by the models carries SynthID, an imperceptible watermark intended to support detection of AI-generated content. Applications that transcode, compress, mix, or otherwise process the output should test whether the watermark survives their audio pipeline.

The release raises the baseline for live voice applications by combining conversation, vision, reasoning, and tool execution in one session. Its production value will depend on the details Google has yet to publish, especially latency, pricing, session limits, and reliability under interruptions.