Gemini 3.8 Live Placed Fifth on the Speech Index, and Extended Thinking Is Now Second

Google shipped Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026 and led with a number one ranking on the Artificial Analysis Speech to Speech Index (82.6). That score belongs to Extended Thinking. The default Gemini 3.8 Live scores 76.0 and sits fifth. When I checked the leaderboard on October 9, Extended Thinking had slipped to second, behind OpenAI's GPT-Live-1 (Astra, medium) at 83.2.

What did Google actually ship on September 15?

Two audio to audio models, one price sheet. Gemini 3.8 Live is the default. Gemini 3.8 Live Extended Thinking adds background reasoning. Google's changelog entry for September 15, 2026 labels both generally available, and they differ in capability contracts, function calling rules and task completion scores.

The model page calls Gemini 3.8 Live the default option for most low latency voice agents. The Extended Thinking page recommends that model when higher background reasoning is required. Both take text, images, audio and video in and return text and audio, with a 131,072 token input limit and a 65,536 token output limit.

The announcement names Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents as integrations, and Salesforce, Genspark and Lumeris as partners. If you already run a managed voice platform, check whether it exposes these model strings yet, because that decides how much of this post you handle yourself.

Which Gemini 3.8 Live model won the number one spot?

Extended Thinking won it on launch day, and the gap between the two siblings is far wider than the 0.6 points that now separate Extended Thinking from the model ahead of it. Google's announcement does not say this in words. I read the values off the Artificial Analysis leaderboard as fetched on October 9, 2026, not from Google's chart images.

Artificial Analysis defines the index as the equal weighted mean of four components: Big Bench Audio reasoning, τ-Voice task completion, an arena preference score and an arena task success rate. I recomputed it from the component values in the page data. Artificial Analysis freezes the arena score when a model is first published, so I used the published component and not today's Elo. The mean uses unrounded components, which is why summing the rounded values shown gives 76.05 for the base model and not 76.0.

On September 15 Artificial Analysis had Extended Thinking ahead of Astra on τ-Voice, 68.6% to 67.9%, which is what I reported on September 16. On October 9 the leaderboard shows Astra at 74.5%, 5.9 points ahead, and 0.6 points ahead on the index. The τ-Voice chart footnotes that both GPT-Live-1 entries rest on a single trial, so I would treat that lead as provisional. Google's number one claim matched the data on launch day and no longer does.

The Gemini 3.1 Flash Live high entry also moved a little since my September 16 table (71.5 then, 72.0 now), which is why every figure here is dated October 9. The sibling gap is the part that has held. The Extended Thinking entry is the High setting, and the docs require a thinking_config there (low, medium or high; minimal is not supported), so the 82.6 and the $3.50 cost below belong to that configuration. The base Gemini 3.8 Live sits 6.6 points below Extended Thinking on the index and 38.5 points below it on τ-Voice, the benchmark that measures whether a voice agent finishes a customer service task. Extended Thinking completes 2.3 times as many of those tasks as the model Google calls the default. Google's post is accurate on the numbers I could check (82.6, 68.6% and 97.7% match the leaderboard data); I could not find its 35.1% Sierra banking figure published anywhere I could open. The framing invites you to carry the flagship number over to the cheaper model, and the data says you cannot.

Is Gemini 3.8 Live better than the model it replaces?

It depends on which thinking level of the old model you were running, a distinction the September 16 version of this post did not make. Google's capabilities guide says gemini-3.1-flash-live-preview defaults to minimal thinking to get the lowest latency. Artificial Analysis lists that default and the high setting as separate entries, and they tell different stories.

Against the high setting, Gemini 3.8 Live gains 4.0 points on the index but loses 7.6 points on τ-Voice and 4.9 on Big Bench Audio, so it is weaker at agentic tasks and audio reasoning than the preview model was when allowed to think. Against the default minimal setting, which is what you get if you never touched the knob, it is better on all three benchmark scores: 11.8 points on the index, 3.9 on τ-Voice and 20.4 on Big Bench Audio. It does not win on speed: 1.18 seconds to first audio against 0.96 for minimal, though both beat the 2.99 seconds of the high setting. In the human preference arena the minimal setting is first and 3.8 Live second, with overlapping 95% intervals, so I would not read that gap as real.

If you ran 3.1 on minimal, you get a quality gain (+11.8 on the index, +3.9 on τ-Voice) for about a fifth of a second. If you ran it on high because your agent had to reason, look at Extended Thinking.

What breaks when you migrate to Gemini 3.8 Live?

Several behaviors change by default and setting proactive_audio: false now returns an error, so a model string swap changes defaults and you should rerun your function calling and video tests. Google's thinking guide says the turn lifecycle stays identical when you move from 3.1 to 3.8 Live. The table below is what the model page adds.

The function calling change is the one I would test hardest. With NON_BLOCKING the model can keep talking while your tool runs, which is what the launch post describes. If a tool result changes what the agent may say, such as "that account is locked", an agent that already said something friendlier has a problem. Blocking mode still exists on the base model through behavior: BLOCKING, and I would keep it on any tool whose result gates the reply. Function calling semantics are the part of a model swap that needs a test call; for another voice model change that takes more than a config edit, see why Eleven v4 Turbo is not a drop in swap.

That escape hatch does not exist on Extended Thinking. There, function calling is async only, blocking mode returns a hard error, and function scheduling is unsupported. The more capable model is also the one that removes your ability to serialize a tool call.

Google's developer channel has a short demo by Thor Schaeff of Google DeepMind covering async tool calls, proactive audio and the switch to Extended Thinking.

What's new in the Gemini Live API

The capability table matters too. Both models support function calling, search grounding, audio generation and thinking. Neither supports context caching, structured outputs, code execution, file search, URL context, Maps grounding or the Batch API. So you cannot cache the long system prompt and knowledge block you will inevitably write, and there is no schema guarantee on the way out; my read is that every structured result has to come back through a function call.

Why does turnComplete no longer mean the turn is complete on Extended Thinking?

Because Extended Thinking keeps reasoning after it stops speaking, so Google changed what an existing protocol field means and added a second field to carry the session state. A working client can break silently when you adopt that model. The docs are direct: turnComplete: true no longer indicates an idle session.

On the base model Google's thinking guide says turnComplete: true reliably signals an idle session, so a client built for it can treat the flag as the end of the exchange: re enable the microphone, close the span, write the transcript row, hand off to the next node in your graph. Under asynchronous reasoning the server may still be running background reasoning or waiting on async tool calls when that flag arrives, and more audio frames or tool calls can follow it.

The replacement is a field called interaction_status with two values. IN_PROGRESS means the server is still working and more output may follow. IDLE means it has finished and is waiting for the user. If you move to Extended Thinking without reading that field, your agent may look correct in testing and then treat the session as idle too early whenever a tool call runs slow. It is a client state machine change that arrives with a model upgrade.

What does Gemini 3.8 Live cost to run?

Audio input runs $0.005 per minute and audio output $0.018 per minute on the paid tier, which prices a three minute support call at about 4 cents and a ten minute call at about 14 cents, and the pricing table charges Gemini 3.8 Live, Extended Thinking and the legacy 3.1 Flash Live Preview at exactly the same rates. I read the pricing page on October 9, 2026 and nothing has changed since launch. Thinking tokens bill at the output rate, $12.00 per million for audio and $4.50 per million for text, so the reasoning premium shows up as volume on your invoice rather than as a higher rate.

Artificial Analysis measured the cost, and my read is that the volume explains the gap. On its fixed 40 question Big Bench Audio subset, completing the test cost $0.84 per hour of input audio for Gemini 3.8 Live and $3.50 for Extended Thinking, a ratio of 4.2 to one at identical list prices. That is a test setting and not a phone call, so it shows direction only. The Batch reported the same two figures on October 2, 2026.

The per call arithmetic below assumes the caller speaks for the whole call and the agent talks for half of it, using the per minute prices Google publishes. The formula is: cost per minute = $0.005 + 0.5 × $0.018 = $0.014, and with video streaming on, add $0.002 for each minute of video. These figures assume the base model. Artificial Analysis's test measured Extended Thinking at 4.2 times the cost per hour of input audio (see below), so do not apply them to it.

Two things sit in that last column. Video input bills at $0.002 per minute, and the new default turn coverage includes any video frames your client sends, which adds a flat 14.3% on top of the audio bill when you stream video. A client that sends no video pays nothing extra. Google's migration note says to send frames only when needed to manage context and cost, so the default is the expensive setting. The same pattern runs through my breakdown of what voice agents cost per deployment.

Where do Google's own docs contradict each other?

Five places, each findable by reading the pages side by side, and each can cost an afternoon, because the pages were updated on different days. Last updated: the model pages September 15, the capabilities guide September 18, the overview October 6.

First, affective dialogue. The model page says it is removed from the API and you must delete enable_affective_dialog. The capabilities guide still documents it with Python and JavaScript samples, noting only that it is unsupported on Gemini 3.1 Flash Live. Second, proactive audio: the model page says it is permanently enabled and that setting it to false returns an error, while the capabilities guide presents it as an opt in you configure in the setup message.

Third, language counts. The overview says 70 supported languages, the launch post says 97, and the capabilities guide, updated after the launch post, now says 99 and lists them. Test the specific language you need rather than trusting any one of those numbers.

Fourth, status. The model pages say Stable, the changelog says generally available, and Google Cloud's Gemini Enterprise Agent Platform release notes say Gemini 3.8 Live went generally available there on September 24, 2026. The capabilities guide still opens with "Preview: The Live API is in preview." The model is stable and the API you reach it through still wears a preview label.

Fifth, and new since September 16: whether Gemini 3.8 Live speaks while a tool runs. The model page makes NON_BLOCKING the default, but the thinking guide's comparison table says of 3.8 Live that the model waits for tool execution before speaking, and credits intermediate filler speech to Extended Thinking. I read that as the base model running tools in the background without narrating them. That is my interpretation, and exactly the behavior to confirm with a test call.

Dates and deadlines since launch week

Three dated events matter if you build on this. On September 24, 2026 Google announced Gemini 3.8 Live with Live Avatar, a video avatar on the same model, available in Gemini Enterprise with custom avatars only through enterprise allowlisting. The same day Gemini 3.8 Live went generally available on Google's enterprise platform. And the deprecations page, last updated October 7, lists gemini-3.1-flash-live-preview with an earliest possible shutdown date of November 17, 2026 and gemini-3.8-live as the recommended replacement. It gives the same date to gemini-2.5-flash-native-audio-preview-12-2025.

That date turns an optional upgrade into a scheduled one, and the migration table above is the work list. The page names the base model as the replacement, so if you ran 3.1 at high thinking for reasoning, my inference (not Google's statement) is that Extended Thinking is the closer match.

What are the hard limits nobody mentions in the launch post?

Session duration is the limit I would plan around first, because the Live API caps an audio only session at 15 minutes and an audio plus video session at 2 minutes, and neither number appears in the announcement. Both caps can be extended with session management.

Two minutes is short for a video support call. A 30 minute call spans two audio sessions or fifteen audio plus video sessions, so plan for one or fourteen reconnects. Google documents session management for extending these caps, so build that before you build features.

The session context window is 128k tokens for native audio output models and 32k for the rest, separate from the 131,072 token model input limit. Google's guide does not say which of the two figures applies to 3.8 Live, so check before you size a long call with video frames. Client authentication is server to server unless you use ephemeral tokens, so a browser client that talks directly to the Live API needs that token exchange built before launch.

What would I actually ship on this?

I would run Extended Thinking where the agent has to follow policy through several tool calls, and the base model where the job is a conversation with short turns and quick tools, and I would test both on my own call recordings before committing. Google's thinking guide draws the same line: the base model for fast tools, Extended Thinking for tools that take seconds and multi step planning.

On September 16 I called the base model's 30.1% on τ-Voice disqualifying for booking and billing work. The arena data makes me back off that. In the Artificial Analysis Speech Agent Arena, where people hold blind voice conversations on scenarios such as booking a dental appointment, Gemini 3.8 Live made the correct final tool calls in 204 of 219 eligible conversations (93.2%, 95% interval 89.0 to 95.8) and Extended Thinking in 229 of 257 (89.1%, 84.7 to 92.4). Those intervals overlap, so the honest reading is that the two are not clearly different on that test, and that the base model is not collapsing on simple booking work. Extended Thinking also ranks tenth by human preference there and the base model second. τ-Voice is a harder test: the paper from Sierra describes 278 customer service tasks across retail, airline and telecom, and Artificial Analysis's implementation judges each against a single correct database end state while the agent follows a policy document.

The two tests describe different jobs. A short appointment booking is something the base model finishes most of the time in the arena. A retail charge dispute that needs a policy lookup and several tool calls is the kind of task τ-Voice scores, and there the base model reaches 30.1% against 68.6% for Extended Thinking. I have not run either model on a client workload for this post, so the split comes from published benchmarks and Google's docs, not my own call logs. What I would not do is take the 82.6 headline, point it at gemini-3.8-live because the docs call that the default, and find out in production that it finishes fewer than one policy heavy task in three.

Check which variant earned a vendor's benchmark number, at which thinking level, on which benchmark and on which date, because a leaderboard can move after you read it. I would measure task completion on your own calls before judging how natural the voice sounds.

If you are working out whether a voice agent belongs in your operation before you argue about model strings, the voice agent pages by industry cover the use cases I take on, and the inbound voice agent page lays out what production work involves. I sell this work, so weigh my opinion accordingly and check the numbers yourself.

Frequently asked questions

Is Gemini 3.8 Live the number one speech to speech model?

No. Gemini 3.8 Live Extended Thinking led the Artificial Analysis Speech to Speech Index at launch with 82.6, but on October 9, 2026 GPT-Live-1 (Astra, medium) is ahead at 83.2. The base Gemini 3.8 Live scores 76.0, which places it fifth, behind two GPT-Live-1 entries, Extended Thinking and Grok Voice Think Fast 2.0 High.

Should I migrate from gemini-3.1-flash-live-preview?

Yes, and before November 17, 2026, the earliest shutdown date Google's deprecations page lists for it (Google gives the exact date on advance notice). Pick the target on purpose. If you ran minimal thinking, Gemini 3.8 Live improves all three benchmark scores. If you ran high thinking for reasoning, the base model is weaker on τ-Voice and Big Bench Audio, so test Extended Thinking.

What does Gemini 3.8 Live cost?

On the paid tier, audio input is $3.00 per million tokens or $0.005 per minute, and audio output is $12.00 per million tokens or $0.018 per minute. Text input is $0.75 and text output $4.50 per million. Gemini 3.8 Live, Extended Thinking and 3.1 Flash Live Preview share these rates, and thinking tokens bill as output.

Why does my Live API client stop early after upgrading?

Most likely you are treating turnComplete: true as the end of the turn. On Extended Thinking that flag no longer means the session is idle, because background reasoning and async tool calls can continue afterwards. Read the interaction_status field on each server message and treat IDLE as the signal that the session is waiting for the user.

How long can a Gemini Live API session run?

Audio only sessions are capped at 15 minutes and audio plus video sessions at 2 minutes. Google's capabilities guide says both can be extended with session management techniques, but the caps apply by default and the launch announcement does not mention them. Plan for session resumption from the first build.