Gemini 3.8 Live

What it is

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are Google DeepMind’s live dialogue models, released on 15 September 2026 and bylined by Tom Ouyang, a principal engineer, and Malini Jaganathan, writing on behalf of the Gemini Audio Team. They are audio-to-audio models: speech goes in, speech comes out, with images, video and text accepted alongside. Google calls them “our most advanced live dialogue models yet”.

The lineage is stated in one line on the model card. “Gemini 3.8 Audio is based on Gemini 3 Pro”, covering both variants, which arrive thirteen days after Gemini 3.8 Flash on the text side of the same generation.

How the pair divides the work

Google’s own split is short. Gemini 3.8 Live is “built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding”. Gemini 3.8 Live Extended Thinking is “built for high-complexity tasks, with increased intelligence and multi-step reasoning”. The developer documentation puts the same division as a routing rule: 3.8 Live is “the default option for most low-latency voice agent experiences”, and Extended Thinking is “recommended when higher background reasoning is required”.

Both are marked stable on the model list, with September 2026 as the latest update, and both support search grounding. Caching, code execution, structured outputs and the Batch API sit outside what the Live API does.

What a session costs

One pricing row covers 3.8 Live, 3.8 Live Extended Thinking and the older gemini-3.1-flash-live-preview, so the reasoning variant is charged at the same rate as the plain one and shows its extra work as tokens consumed.

There is a free tier on the API, with the usual data difference: the pricing page records free-tier prompts as used to improve Google’s products, and paid-tier prompts as excluded from that. Search grounding gives 5,000 free requests a month shared across all Gemini 3.x models, then $14 per 1,000.

Artificial Analysis measures the practical version of this on its own fixed audio subset, where an hour of input audio works out at $0.84 through 3.8 Live and $3.50 through Extended Thinking, read on 15 September 2026.

Where it lands on the independent board

Google’s launch claim is that Extended Thinking takes “the #1 overall spot on Artificial Analysis’ Speech to Speech Quality Index (82.6)”. Artificial Analysis’s own speech board, read on 15 September 2026, carries the same figure at the top of the index.

That index is a composite, and the components split differently. On Big Bench Audio, Extended Thinking scores 97.7%, the figure Google published and the one Artificial Analysis records, which places it fourth on that single test behind StepFun’s StepAudio 3 Realtime at 99.7%, Alibaba Cloud’s Qwen Audio 3.0 Realtime Plus at 99.2% and Qwen3.5 Omni Plus Realtime at 98.7%. Plain 3.8 Live scores 91.7% on the same test. The top placing belongs to the composite, and the reasoning row underneath it is tighter than the headline reads.

Agentic work is where the lead is widest. Google reports 68.6% on τ-Voice, and the same 68.6% average sits first on Artificial Analysis’s board, ahead of GPT-Live-1 (Astra, medium) at 67.9%, with plain 3.8 Live at 30.1%. Google also reports 35.1% on Sierra’s τ-Voice-banking benchmark, which is the company’s own figure on its own launch day.

On latency, Artificial Analysis measures time to first audio at 1.18 seconds for 3.8 Live and 1.35 seconds for Extended Thinking, the cost of reasoning before the first word.

Google says 3.8 Live secured “a second place in the Speech Agent Arena”, a human preference board rather than a test set, and Artificial Analysis’s arena Elo agrees: 3.8 Live sits at 1,083, second behind Google’s own Gemini 3.1 Flash Live Minimal at 1,096, with Extended Thinking down at 990. Readers preferring the cheap fast model to the deliberating one is what that ordering describes, which is a point about what live conversation rewards.

The remaining launch benchmark is ServiceNow’s EVA-Bench, where Google says its models “push the Pareto Frontier for complex workflows by successfully balancing accuracy with conversational quality”. Google’s own footnote to that chart states the run happened “on the Live API on Gemini Enterprise Agent Platform”, and the post publishes the claim without the underlying scores.

Speaking and reasoning at the same time

The interesting engineering is in how Extended Thinking keeps talking while it works. Google describes a model that “reasons and speaks simultaneously”, using early verbal cues such as Let me check that… to acknowledge a request, then “live progress narration to walk users through multi-step background tasks as they progress”. Both models execute tools and API calls in the background while the conversation continues.

For anyone building on it, the documentation turns that into a protocol change worth reading before the upgrade. On Extended Thinking, turnComplete: true now means the turn finished rather than the server being idle, and the client reads the interaction_status field instead: IN_PROGRESS while background reasoning or asynchronous tool calls are still running, IDLE once everything has finished. Function calling is non-blocking only, and synchronous blocking mode “returns a hard error”.

The plain 3.8 Live has its own changes from gemini-3.1-flash-live-preview. Asynchronous execution is now the default calling mode, thinking_level is unsupported, proactive audio is permanently enabled so setting proactive_audio: false returns an error, affective dialogue is removed from the API, and turn coverage defaults to TURN_INCLUDES_AUDIO_ACTIVITY_AND_ALL_VIDEO, which sends video frames by default. Google flags that last default as a context and cost consideration, and advises sending frames only when they are needed.

The language handling is the consumer-facing headline: the model “automatically detects and transitions between 97 supported languages mid-conversation”, on Google’s own description, so a speaker switching language part-way through a sentence stays in the same session.

Where you can use it today

Both models began rolling out on 15 September 2026.

• Developers: the Gemini API and Google AI Studio, for both models.
• Enterprises: private preview in Gemini Enterprise, with Gemini Enterprise for Customer Experience listed as coming soon, joined for Extended Thinking by Google Workspace business customers.
• Everyone: 3.8 Live in Search Live; Extended Thinking in Gemini Live, in Workspace Docs for Google AI Pro and Ultra subscribers, and in Gmail and Keep for all Google AI subscribers.

Google names Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents as the developer platforms carrying the Live API, and Salesforce, Genspark and Lumeris as customers it says are “excited about” the pair, citing “impressive latency, fluidity, and tool-calling capabilities”.

Watermarking and the small print

Every piece of audio these models generate is watermarked. Google’s line: “This imperceptible watermark is woven directly into the audio output, ensuring AI-generated content remains detectable to help prevent misinformation.” SynthID covers all audio from Google’s AI products, so the check applies to output from the consumer surfaces as well as the API.

The model card states a knowledge cutoff of January 2025, inputs of “audio, images, video, and text with a token context window of up to 128K”, and outputs of “audio and text, with 64K token output”.

On safety, the card reasons across models rather than from a fresh run: “We evaluated Gemini 3.7 Flash as outlined in our latest Frontier Safety Framework, and we found that it did not reach any Tracked or Critical Capability Levels (T/CCLs)”, concluding that “based on Gemini 3.7 Flash results, we are confident that Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are not likely to reach any T/CCLs”. So the published assurance for the audio pair rests on the evaluation of Gemini 3.7 Flash, a text model two releases back. Google’s stated limitations are hallucinations, “occasional slowness or timeout issues”, and continuing work on jailbreak resistance.

Related

• Gemini 3.8 Flash, the same generation on the text side
• Gemini 3.7 Flash, the model whose safety evaluation this pair is assured against
• The Gemini 3.5 family, for the family picture