On September 15, 2026, Google DeepMind added two new models — Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking — to its model-cards channel, both published under a single "Gemini 3.8 Audio" umbrella card. The story is not new parameters and not new benchmark numbers. The story is that Google has merged two previously separate tracks — real-time conversation and audible reasoning — into a single audio output stream.

Not a new architecture, just Gemini 3 Pro with native audio unlocked

Let me get the architecture out of the way first. Neither model is trained from scratch. The model card states plainly that "Gemini 3.8 Audio is based on Gemini 3 Pro," and points readers at the Gemini 3 Pro card for training data, training processing, hardware and software details. The "Live" and "Live Extended Thinking" labels essentially mean that Google has promoted "audio" from a supported input modality to a native output modality on top of Gemini 3 Pro's native multimodal stack. The model no longer hands text to an external TTS — it generates speech tokens end-to-end.

On the input side, both models accept 128K-token context, with modalities covering audio, image, video and text (16-bit PCM audio at 16 kHz). On the output side, the model emits text plus audio together, with 64K tokens of text headroom and 16-bit PCM audio at 24 kHz. Architecturally the model is a 40-layer Transformer, but the model card explicitly defers backbone details to the Gemini 3 Pro card. There is no paper-level structural innovation here — the real engineering work is in streaming multimodal token scheduling and in the latency optimization that makes "generate while speaking" actually feel real-time.

Two tiers: cheap and fast, or think-out-loud

Same backbone, but the two models are clearly split. DeepMind's official audio page states the distinction bluntly:

  • Gemini 3.8 Live is positioned for "high-throughput, cost-sensitive, near real-time voice interfaces" — the typical scenario is customer-service bots, e-commerce shopping assistants, NPC companions in games, anything where the conversation itself does not need complex reasoning but must respond in under a second.
  • Gemini 3.8 Live Extended Thinking is positioned for "complex reasoning and high-complexity tasks" — its signature move is to "think out loud": before it starts speaking, the model enters a thinking interval, and then narrates its own reasoning progress in real time with a natural voice, instead of thinking silently and only at the very end pushing a TTS-read summary.

The second tier is the most interesting product decision here. Almost every "thinking" AI assistant today — whether OpenAI's o-series or Anthropic's thinking mode — has the same experience break: the model sits silent for tens of seconds, then suddenly starts talking. From the user's perspective, there is no feedback signal that says "the model is thinking" versus "the model is stuck." Extended Thinking turns that silent thinking interval into audible thinking. The user can hear the model mumble to itself, watching it change direction in real time. This "process-visible reasoning" has not been done on the audio side in OpenAI's or Anthropic's current flagship products. Google is the first to drag thinking mode into the speech stream itself.

Distribution: from Gemini App to Gmail and Docs

The model card lists Google-as-distributor across the board:

  • For 3.8 Live: Gemini API, Gemini App, Google AI Studio, Google Cloud / Vertex AI, and Google Search Live.
  • For 3.8 Live Extended Thinking: the same four plus Google Workspace (Gmail, Docs, Keep).

That extra Workspace integration is the real landing signal — Extended Thinking is not just for developers building agents. Gmail, Docs and Keep are the three Workspace apps users open twenty times a day. Google is preparing to push "thinking-out-loud voice assistants" into the daily email and document workflow, not just AI Studio demos.

The integration bar for developers has been deliberately lowered. The Live API docs page (ai.google.dev/gemini-api/docs/live-api, last updated 2026-09-10) makes WebSocket duplex streaming the default protocol and ships full sample code for the GenAI SDK, raw WebSocket, and the Agent Development Kit (ADK) Streaming path. On the third-party side, LiveKit, Pipecat, Fishjam, Voximplant, Agora, and Firebase AI Logic are all in the partner list. For real-time voice-agent developers, this distribution density plus Google's own Search and AI Studio surface means switching to Gemini 3.8 Live is not a rewrite — it is a one-line model-id swap on top of an existing WebRTC / WebSocket real-time stack.

Evaluation and safety: Sierra τ³-Bench makes it into the official methodology

The model card stays restrained on evaluation, and points to a dedicated Evals & Methodology subpage (deepmind.google/models/evals-methodology/gemini-3-8-live). The actual benchmarks are run by three independent parties:

  • ServiceNow EVA-Bench — a multi-turn bot-to-bot spoken conversation framework, run on the Gemini Enterprise Agent Platform with both thinking-minimal and thinking-high configurations.
  • Artificial Analysis — an independent third-party benchmark that measures speech-to-speech models on reasoning quality, conversational dynamics, generation latency, and price.
  • Sierra τ³-Bench — tests whether agents can navigate a large unstructured knowledge base and execute multi-step tool calls to resolve realistic banking workflows. The runs use the Gemini API with model-ids gemini-3.8-live-preview and gemini-3.8-live-extended-thinking, under thinking-high and default sampling.

On safety, every audio output is automatically tagged with SynthID, Google's own audio watermarking technology. The Frontier Safety assessment carries forward Gemini 3.7 Flash's conclusion: 3.8 Live and Extended Thinking do not represent a material capability jump over 3.7 Flash, so neither model triggers any Tracked or Critical Capability Level. The knowledge cutoff is January 2025.

Industry meaning: "reasoning in the audio stream" as a product feature

Zoom out: from the second half of 2026 onward, every frontier model lab has moved differentiation away from "who is longer" and toward "who can hold steady inside a multimodal stream." GPT-Realtime 2.1, Gemini 3.1 Flash TTS, Inworld TTS-2, Cartesia Sonic 3.6 — this parade of real-time audio and speech models shipped in May through September, and it tells you the real-time voice track has already moved past "can it be done" and is now squarely in the "how do you make it stop making the user wait" phase.

Gemini 3.8 Live Extended Thinking takes a different direction in that race — it does not chase latency numbers, it chases "visible reasoning." Out-loud chain-of-thought will increase the per-turn response time (the user has to listen to the model talk to itself), but over the longer arc it relocates "what is the model doing" from developer logs into the end-user's ear. If this paradigm is validated, OpenAI and Anthropic following suit is almost certain. When that happens, "thinking-out-loud speech models" will likely become the default form of the next generation of voice assistants.

For Chinese vendors, the comparison is direct. Today the domestic real-time speech models — Alibaba CosyVoice, ByteDance Seed-TTS, Tencent Hunyuan TTS, and so on — are mostly optimizing on "voice quality + real-time ratio." None of them have pulled thinking mode end-to-end into the speech stream. If any of them ships "think-while-speak" before the end of Q4 2026, they will get to eat the product-level dividend that this Google pattern sets up.

So what

The Gemini 3.8 Audio release is not about a new model — it is Gemini 3 Pro with a native audio output. The substance is the combination of three things: native audio output, visible thinking, and a Workspace-wide distribution channel. For the first time, Google packages "real-time voice + reasoning interpretability + consumer-grade distribution" as a single product surface. Developers can hit the API directly, and ordinary users will soon hear what their model is actually thinking inside Gmail and Docs. This is the marker that the real-time voice-agent track is moving from demo to daily tool.