Trend Watch 007 | Google Releases Gemini 3.8 Live and Extended Thinking: A Voice Agent That 'Thinks Out Loud', and a Self-Proctored Exam Paper

Trend Watch: hot releases × technical judgment × practical advice. Author: Yongliang

---

On September 15, Google’s official blog (signed by Tom Ouyang and Malini Jaganathan of the Gemini Audio Team) released two voice models at once: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The former emphasizes scale and cost efficiency, highlighting conversational intelligence and visual understanding; the latter targets highly complex tasks, with a selling point of “reasoning and speaking at the same time” — the model thinks while it talks, and when it needs to verify something it first says “Let me check that…”, then reports its progress in real time. That same day, developers could call the models through the Gemini API and AI Studio; individual users will see the new features roll out in batches across Search Live, the Gemini Live app, and Workspace; the enterprise edition remains in private preview. The voice agent space has been hot this year, with every player talking about “getting real work done” rather than just “chatting”. Google has gone all-in with its framing this time: per the official account, Extended Thinking took first place (82.6) on Artificial Analysis’s Speech to Speech Quality Index, a third-party leaderboard. But this scorecard deserves a line-by-line reading — because the proctor, the test-taker, and the one reporting the scores are, in no small part, the same company.

What Was Released

First, the hard facts as officially stated, all from the September 15 blog post:

The two models have a clear division of labor. Gemini 3.8 Live takes the scale route: cost efficiency first, with automatic in-conversation switching across 97 languages, and it can accept near-real-time visual input — the launch demos included reading a chess position and writing a React component directly from a sketch. It can also run tool calls and API requests in the background without interrupting the current conversation. Gemini 3.8 Live Extended Thinking takes the high-complexity route: it handles multi-step reasoning tasks, and its distinguishing feature is “speaking the thought process out loud” — users can hear when the model is looking something up and when it is computing, instead of staring at silence waiting for a result.

Availability comes in three tiers. The developer tier went live on day one: the Gemini API and AI Studio opened in sync. The consumer tier is tiered by subscription: Search Live, the Gemini Live app, and Docs Live / Gmail Live / Keep Live in Workspace follow the Google AI Pro and Ultra subscription tiers — not everyone got all the features on the same day. The enterprise tier is still in private preview: Gemini Enterprise customers have to apply.

The supporting materials are complete too: audio output carries a SynthID watermark, and an official model card has been published. The partner ecosystem list runs long — Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, Vision Agents; customer endorsements came from Salesforce, Genspark, and Lumeris.

How to Read the Official Scorecard

This section demands a slow read, because the numbers come from different sources and carry different weight.

Two come from third-party leaderboards: Extended Thinking ranks first on Artificial Analysis’s Speech to Speech Quality Index with a score of 82.6, and second on Speech Agent Arena’s user-preference leaderboard. These are tests by external organizations and can count as independent evaluations.

Then there are the numbers from the company’s own account: a 68.6% agent task completion rate on τ-Voice, 35.1% on the Sierra τ-Voice-banking scenario, and 97.7% on Big Bench Audio. None of these figures has been independently reproduced by a third party so far, so in this piece they can only be called “official claims”.

The most interesting item is ServiceNow’s EVA-Bench. The blog post says Extended Thinking “pushes the Pareto frontier of accuracy and conversational quality”, but the claim is immediately followed by a footnote of its own: the tests were run on the Live API of Gemini Enterprise Agent Platform. In plain terms — this enterprise customer-service benchmark ran on Google’s own agent platform. EVA-Bench itself is indeed ServiceNow’s evaluation framework; that part is true. But the runtime environment was the company’s own stack. That caveat was written by the company itself, and readers should remember it for themselves.

So the way to read this scorecard is: the third-party portion can be cited, the self-reported portion needs a label attached, and the EVA-Bench item requires holding both halves in mind — “third-party framework” and “home track”.

The Sober Other Half

This is the part this issue really wants to talk about seriously. Five points, ranked by weight.

First: a self-hosted launch plus self-run benchmarks — honest, but not entirely honest. It is industry practice for AI vendors to publish their own results on their own platforms, so no surprise there. But the honesty material this time was ready-made: the EVA-Bench footnote — “run on the Live API of Gemini Enterprise Agent Platform” — amounts to the company flagging its own conflict of interest. This is not cheating — the caveat is stated out in the open; but neither is it independent verification. The real gap is this: nowhere in the launch materials has a single external organization reproduced the pretty numbers like τ-Voice and Big Bench Audio. The launch buzz will fade; whether any third party re-tests these numbers is what will decide whether they hold up.

Second: the naming of “Extended Thinking” — is thinking out loud reasoning transparency, or anthropomorphic theater? With text models, a chain of thought can be audited word for word, screenshotted, challenged. Voice is different: what you hear is a smooth “Let me check that…”, but is the model actually retrieving and computing, or merely soothing you with a trained verbal routine? The “thinking” in voice interactions cannot be audited the way text can — the listener has neither access to the reasoning process nor any way to tell “genuine multi-step reasoning” from “performed multi-step reasoning”. Wrapping latency up as “thinking” is, in engineering terms, an elegant product decision; in narrative terms, it quietly walks off with the trust that the term “extended thinking” accumulated in the text era. Between transparency and performance lies a layer of voice — and for now, the vendor defines what that layer means.

Third: between promise and delivery sit a subscription table and a preview waitlist. On launch day, developers could use the API — that part is true. But on the consumer side, the full Workspace suite — Docs Live, Gmail Live, Keep Live — is tiered by Pro / Ultra subscription; between “it launched” and “you can actually use it” stands a paywall. On the enterprise side, the core selling point, Gemini Enterprise, is still in private preview — application-based, opaque, no timeline. This is not a Google-only problem; “announce everything at the launch event, deliver in tiers” is the industry norm. But for readers, distinguishing the three tiers of “launched” matters far more than remembering the demo videos from the keynote.

Fourth: the classic trio of voice-agent metrics once again has no independent answers. Judging whether a voice agent can actually do work comes down to three metrics you cannot avoid: interruption handling (how quickly the model stops when a user cuts in), end-to-end latency (how long you wait before it starts speaking), and tool-call reliability (it says it called the API — did the call actually succeed?). In this launch’s materials, the company provided no third-party reproduction data for any of these three metrics. The chess demo and the React component demo were impressive, but a demo is a one-time success, while the classic trio is about success rates over ten thousand runs. Until independent evaluations arrive, take “can do real work” with a grain of salt.

Fifth: SynthID is a detection watermark, not an accountability solution. Watermarking AI audio is a good thing; it makes later identification easier. But a watermark answers the question “was this audio AI-generated?” — it cannot answer “who is responsible when this AI audio says something wrong?”. By the time a voice agent has looked up the material for the user, read out the conclusion, and finished its background tool calls, the chain of responsibility when something goes wrong — model, platform, integrator, customer — is not covered by any watermark today. Presenting a detection tool as an accountability framework is a common sleight of hand in launch-event narratives.

If You Are a Developer

In three buckets — “usable now / wait and see / what to look at when picking”:

Usable right now: The Gemini API and AI Studio went live on day one, so you can start prototyping voice agents immediately. Automatic switching across 97 languages, if it really performs as officially claimed, is a genuinely valuable selling point for multilingual customer support and products going global — but run a round of testing on your own corpus first, and do not treat the keynote demos as your acceptance criteria. The design of background tool calls that do not interrupt the conversation deserves focused hands-on time in your prototype; it is what separates an “assistant” from an “employee”.

Wait and see: For the Agent Platform capabilities inside Gemini Enterprise’s private preview, hold off evaluating until general availability plus independent evaluations; for the Workspace suite’s subscription tiers, check which tier your team actually falls into before purchasing; and for the “thinking out loud” progress narration, design your own product treating it as a UI cue rather than evidence of reasoning — your users cannot audit a stretch of audio any better than you can.

Metrics to watch when picking: Whichever vendor you finally choose, voice agent selection comes down to five numbers — interruption response time, end-to-end latency, tool-call success rate (broken down by scenario, not aggregated), the real usable rate of multilingual switching (test your target languages, not the languages in official demos), and per-task cost (multi-step reasoning modes like Extended Thinking cost several times more than standard models — get the exact figures). None of these five numbers will be handed to you at a launch event; you will have to benchmark them yourself, or wait for third-party organizations to test.

Closing Thoughts

The technical substance of Google’s release is real: reasoning and speaking at the same time, background tool calls that do not interrupt the conversation, switching across 97 languages — any one of these alone is enough to make rivals nervous. But the narrative padding at the launch event is real too: a self-proctored exam paper, “thinking” that cannot be audited, an “everything launched” that is tiered by subscription table, the classic trio with no third-party reproduction, and a detection watermark dressed up as an accountability framework. The voice agent race has entered its second half, and what is being compared is no longer whose demo dazzles more, but whose numbers survive outside re-testing. Gemini 3.8 Live has handed in its paper; the grading pen is in third parties’ hands — this issue we have first transcribed the exam sheet. Whether the answers are right, we will talk after someone else has finished grading.

References

• Google official blog: the launch post for Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking (2026-09-15, signed by Tom Ouyang and Malini Jaganathan, Gemini Audio Team)
• The model card accompanying the blog post (2026-09-15)
• Third-party leaderboards cited in the blog post: Artificial Analysis Speech to Speech Quality Index, Speech Agent Arena (rankings as of 2026-09-15)