Google's Gemini 3.8 Live Reasons Out Loud While It Talks

Google introduced two new conversational models on September 15, 2026, and the more interesting one does something previous voice assistants could not do cleanly: it keeps reasoning after it has started answering. Gemini 3.8 Live Extended Thinking talks and works through a problem at the same time, narrating its progress instead of going silent while it thinks. The companion model, Gemini 3.8 Live, is the cheaper, faster option aimed at the voice agents that companies are now building at scale.

Related reading: Google's Gemini 4 Argon Reaches Trusted Cyber Defenders First, With a 1 Million-Token Output Limit · DeepSeek's Cheaper V4.1-Flash Beats Its Own Pro Model on Agent Tasks.

The launch, announced on the Google blog, is a product story rather than a research paper. Google is positioning both models as the plumbing for production voice systems, sold through the Gemini Live API and surfaced in the Gemini app, Google Workspace, and Search. The company also leaned heavily on third-party benchmarks, which it published alongside the announcement.

What Google shipped

The two models split the usual tradeoff between speed and depth.

• Gemini 3.8 Live is described as built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding.
• Gemini 3.8 Live Extended Thinking targets high-complexity tasks, with higher intelligence and multi-step reasoning. Its distinguishing feature is that it reasons and speaks at once.

The technical details Google shared are mostly about behavior, not architecture. Google says 3.8 Live processes visual input in near real time, so a camera feed or a shared screen becomes context the model can refer to mid-sentence. It detects and switches between 97 supported languages during a single conversation. Both models run tool calls and API requests in the background while the spoken exchange continues, which lets them acknowledge a request, keep chatting, and report back when a task finishes.

Extended Thinking handles the awkward pause that plagues voice assistants assigned something hard. Instead of dead air, it uses verbal stubs like "Let me check that" and then narrates its progress as it works. In Google's demos, that ranges from turning a hand sketch plus spoken corrections into a working React component, to coordinating multi-step bookings with asynchronous function calls.

The benchmark numbers

Google's strongest claims rest on the Artificial Analysis Speech to Speech Quality Index, where 3.8 Live Extended Thinking took the number one overall spot at 82.6. That is an independent evaluation, not a Google self-report, which makes it the most load-bearing number in the announcement.

The agentic results are where voice models are usually tested hardest, because they have to complete a task rather than merely sound fluent. Google reports 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking benchmark. The second figure is the more revealing one: banking is full of multi-step, high-stakes procedures, and 35% completion is a reminder that even the best current systems fail most of the time on that class of task. Google also reports 97.7% on Big Bench Audio, which measures perception rather than action.

Gemini 3.8 Live did not top the chart. Google says it took second place in the Speech Agent Arena, a separate comparison, while remaining cheaper to run. That framing is deliberate: Google is selling 3.8 Live as the volume model and Extended Thinking as the premium one, and it wants developers to choose between them by cost rather than write off the family after one billing cycle.

On ServiceNow's EVA-Bench, a voice-agent benchmark for complex workflows, Google says its models push the Pareto frontier by balancing accuracy against conversational quality. That result came from a run on the Live API on the Gemini Enterprise Agent Platform, so it reflects the enterprise configuration rather than a raw model call.

Where you can use it

Rollout began the day of the announcement, and Google split the two models more than a typical release.

Gemini 3.8 Live is available to developers in the Gemini API and Google AI Studio, to enterprises in private preview through Gemini Enterprise, and to everyone in Search Live. Extended Thinking is available in the same developer surfaces and the same enterprise preview, but it also reaches Workspace business customers and, for consumers, Gemini Live, Docs Live, Gmail Live, and Keep Live for Google AI Pro and Ultra subscribers.

The consumer tiering matters. The vision Google is selling is a voice assistant that can act across your documents and mail while you talk, and that capability is gated behind a subscription rather than the free tier. The free version in Search Live is a narrower, mostly informational experience.

The developer stack underneath

A feature like speaking while reasoning only matters if developers can build on it, and Google spent a sizable part of the post listing the platforms that already integrate the Gemini Live API. That group includes Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents, which handle real-time media streaming so application teams do not have to. Google also named Salesforce, Genspark, and Lumeris among partners testing the models, with the usual focus on latency and tool-calling.

The pattern here is familiar from the cloud era: the model vendor provides the model, and a layer of middleware vendors provides the plumbing that turns a model into a product. For enterprises weighing a voice rollout, the maturity of that middleware may matter as much as the benchmark scores, because the hard part of a voice agent is rarely the model itself. It is the turn-taking, interruption handling, and failover when a call drops.

Signals about safety and provenance

Google paired the launch with two safety notes. Every piece of audio produced by its AI products carries a SynthID watermark, an imperceptible signal woven into the output that Google says helps detection tools flag AI-generated content. The company also published a model card for the audio models.

Neither measure is a complete answer to the misuse problem. Watermarks help platforms that look for them, and a determined actor can record and re-encode audio in ways that may weaken the signal. But the disclosure is now standard practice across the major labs, and Google's decision to apply it by default, rather than as an option, is the more significant choice.

Context: a loud week for AI

The release landed in a period of unusual public hand-wringing about AI risk. Days earlier, Dario Amodei of Anthropic published an essay calling on the industry to slow frontier development, and that call drew support from OpenAI's Sam Altman and others. Microsoft separately published a draft code of conduct requiring its models never to resist correction or shutdown.

A voice model that reasons out loud sits a little awkwardly beside that conversation. It is an incremental capability, not a step toward autonomous superintelligence, but it is exactly the kind of model that makes an always-listening assistant more plausible, and plausibility is what shapes public comfort. Google did not address the tension directly. It published a model card and moved on.

Caveats and what to watch

A few things are worth holding separate from the marketing.

• Benchmarks are a snapshot, and some are vendor-selected. Artificial Analysis and ServiceNow runs are independent of Google, but Google chose which to highlight, and agentic benchmarks like τ-Voice and EVA-Bench depend heavily on the harness and the tools provided.
• The jump from demo to deployment is large. Multi-step booking demos are scripted to succeed. Real callers change their minds, mumble, and talk over the agent, and the 35.1% figure on τ-Voice-banking suggests the failure modes are not rare.
• Availability is gated. The most useful capabilities for consumers sit behind Google AI Pro and Ultra, and the enterprise features are in private preview.
• Pricing is the real test. Google's pitch for 3.8 Live is cost per completed task, a metric that only gets settled once developers run it against their own workloads.

The bigger question is whether "reason and speak simultaneously" becomes a baseline expectation or a niche feature. If the latency claims hold up in production, voice agents stop being a novelty layered on text chatbots and become a distinct product category, one where the interface is the conversation and the work happens in the background. Google's release is a bet that the category is real. The next two quarters, when the developer platforms ship their own integrations and the first public cost numbers appear, will show whether users agree.

Cover: Google Headquarters (Googleplex), Mountain View, by Asoundd (CC BY-SA 4.0) via Wikimedia Commons.