Gemini 3.8 Live Review 2026: Google Split Its Voice Line in Two, and the Cheap One Regressed
Gemini 3.8 Live Review 2026: Google Split Its Voice Line in Two, and the Cheap One Regressed
Verdict first: On September 15, 2026, Google shipped two voice models rather than one — Gemini 3.8 Live for cheap, fast, high-volume dialogue, and Gemini 3.8 Live Extended Thinking for multi-step reasoning that happens while the model keeps talking. The headline result is real: Extended Thinking took #1 on Artificial Analysis's Speech to Speech Quality Index at 82.6, ahead of GPT-Live-1 Astra (81.5) and Grok Voice Think Fast 2.0 High (81.3). The problem is what sits underneath. The cheap base model appears to have regressed against its own predecessor — 30.1% on τ-Voice against the older Gemini 3.1 Flash Live's 37.7% — and it ranks fifth on that same quality index, below both GPT-Live-1 configurations. Add the fact that Google published no latency figures at all, and a demo in which the model lost to a basic chess checkmate pattern, and this becomes a release where the flagship tier is genuinely strong and the volume tier is genuinely questionable.
Gemini 3.8 Live & Extended Thinking — one tier leads, the other regressed
What Is Gemini 3.8 Live?
Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are Google's real-time voice models, announced on September 15, 2026 and described by the company as "our most advanced live dialogue models yet." They are native audio-to-audio models — speech in, speech out — rather than a speech-to-text pipeline bolted onto a text model and then back through text-to-speech.
The framing Google uses is worth reading closely. These are not positioned as better chatbots that happen to speak. They are positioned as voice as an interface for work — the model takes a request verbally, goes away and does something, and keeps the conversation alive while it does. That distinction explains almost every design decision in the release, including the ones that create problems.
At a glance:
| Attribute | Gemini 3.8 Live |
|---|---|
| Announced | September 15, 2026 (page updated September 17) |
| Type | Native audio-to-audio, real-time dialogue |
| Models | 3.8 Live (base) and 3.8 Live Extended Thinking |
| Context window | 128K tokens |
| Languages | 97, with mid-conversation switching |
| Visual input | Near-real-time visual grounding |
| Tool use | Runs tools and API calls in the background while conversation continues |
| Audio watermark | SynthID on all audio output |
| Model ID (Extended Thinking) | gemini-3.8-live-extended-thinking |
| Access | Gemini API, Google AI Studio, Vertex AI, private preview in Gemini Enterprise |
| Status | Preview, not general availability |
| Session caps | Audio + video 2 minutes; audio-only 15 minutes |
Two Models, Two Different Jobs
The most important thing to understand about this release is that the two models are not quality tiers of one another. They are built for different jobs, and choosing wrong means either overpaying or underperforming.
| Gemini 3.8 Live | 3.8 Live Extended Thinking | |
|---|---|---|
| Built for | Scale and cost efficiency | High-complexity, multi-step tasks |
| Reasoning style | Direct response | Reasons while speaking — "Let me check that…" |
| Background work | Runs tools while conversation continues | Narrates progress on multi-step work aloud |
| Strengths | Speed, cost, visual grounding, 97-language switching | Deeper reasoning, enterprise task completion |
| Completion signal | turnComplete: true → returns to idle | Track interaction_status (IN_PROGRESS / IDLE) |
| Client complexity | Simple state model | More complex — can speak fillers while reasoning |
| Cost per hour of input audio | ~$0.84 (cheapest on the panel) | ~$3.50 |
That row about the completion signal is the one engineers will feel first. The base model tells you when it is done and goes idle, which is a clean contract for a client. Extended Thinking can be speaking filler while still working, so a client cannot treat "the model is talking" as "the task is running" or "silence" as "the task is done." If you are building on this, that is a real architectural difference, not a detail.
Two further API notes from the release: proactive audio is permanently on in the newer models, and affective dialog was removed from the API. Both are the kind of change that breaks existing integrations quietly.
Benchmarks: A #1 and a Regression
There are two stories in the numbers, and most coverage reported only the first.
The good story: Extended Thinking tops the voice index
Gemini 3.8 Live Extended Thinking took #1 on Artificial Analysis's Speech to Speech Quality Index at 82.6, ahead of GPT-Live-1 Astra at 81.5 and Grok Voice Think Fast 2.0 High at 81.3. Google also reports 68.6% on τ-Voice for agentic task completion, 35.1% on Sierra's τ-Voice-banking, and 97.7% on Big Bench Audio, and says the models push the Pareto frontier on ServiceNow's EVA-Bench for complex workflows.
| Model | Speech to Speech Quality Index |
|---|---|
| Gemini 3.8 Live Extended Thinking | 82.6 (#1) |
| GPT-Live-1 Astra | 81.5 |
| Grok Voice Think Fast 2.0 High | 81.3 |
| GPT-Live-1 (second config) | Above 76.0 |
| Gemini 3.8 Live (base) | 76.0 (#5) |
The bad story: the base model regressed
As reported by DataCamp, the base Gemini 3.8 Live scores 30.1% on τ-Voice, against 37.7% for the older Gemini 3.1 Flash Live. That is a 7.6-point decline on the benchmark that measures whether a voice model can actually complete tasks — and it is a decline relative to Google's own previous generation, not to a competitor.
The base model also places fifth at 76.0 on the quality index, below both GPT-Live-1 configurations. In other words: the tier Google is pitching as the cost-effective default for developers is measurably worse at agentic voice work than the model it replaces. If your use case involves tool calling or task completion rather than conversation, the base model should not be your choice.
What to hold onto about the numbers
- Several figures are vendor-reported and unreproduced. Google cites the Artificial Analysis index (which is independent) but also τ-Voice, τ-Voice-banking and Big Bench Audio (which are its own), and at least one was run inside Gemini Enterprise's platform. Commentators questioned how much the harness contributed to the result.
- Nobody has run the two models against each other independently. One published comparison states this outright, and another notes it did not run the models at all. Most available coverage is analysis of Google's numbers rather than independent testing.
- The only substantial hands-on review available covers the earlier Gemini Live, not 3.8 — so hands-on evidence for this specific release is genuinely thin.
Pricing: The Cheapest Voice Tier on the Panel
Price is where the release is most defensible, at least for the base model.
| Metric | Price |
|---|---|
| Audio input (Live API) | $0.005 per minute |
| Audio output (Live API) | $0.018 per minute |
| Two-way audio, effective | ~$1.38 per hour |
| Input audio, base 3.8 Live (AA measurement) | ~$0.84 per hour — cheapest on the panel |
| Input audio, Extended Thinking (AA measurement) | ~$3.50 per hour |
| Text tokens (Live tier) | $0.75 / 1M in · $4.50 / 1M out |
| Audio tokens (Live tier) | $3.00 / 1M in · $12.00 / 1M out |
| Image / video input | $1.00 / 1M |
Two cautions. First, Google's own announcement published no specific figures — only a promise of a "highly competitive price point" and a chart on cost per hour. That means early claims about value were speculative until third parties measured it.
Second, there is a known data-quality trap in the reporting: an earlier "$1.50 per hour" reading that circulated actually belonged to Gemini 3.1 Flash Live, not to these new models. If you have seen that number attributed to 3.8 Live, it is wrong.
Availability Is Fragmented
Rollout is staggered in a way that will frustrate anyone planning a deployment.
| Surface | Base 3.8 Live | Extended Thinking |
|---|---|---|
| Gemini API | Yes | Yes |
| Google AI Studio | Yes | Yes |
| Vertex AI | Yes | Yes |
| Gemini Enterprise | Private preview | Private preview |
| Gemini Enterprise for CX | Coming soon | Coming soon |
| Search Live | Everyone | — |
| Gemini Live | — | Everyone |
| Workspace Docs | — | Google AI Pro / Ultra subscribers |
| Gmail / Keep | — | All Google AI subscribers |
Note the asymmetry: Search Live gets the base model, Gemini Live gets Extended Thinking. And some Workspace users report being stuck on Gemini 3.6 or 3.7 while seeing seat-price increases — a rollout complaint that predates this release but is relevant to anyone budgeting for it.
Developer platforms building on the Live API include Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents, with Salesforce, Genspark and Lumeris named as partners. That ecosystem breadth is a genuine strength — voice agents are an integration-heavy product, and having this many established platforms supporting the API shortens the path to production considerably.
The Latency Problem (And Why It Matters More in Voice)
Google published no latency figures for either model. No first-audio timing, no p50, no p95. For a release whose entire pitch is real-time conversation, that is a striking omission.
The follow-on problem is that latency in voice is not a performance metric — it is a product-quality metric. As one analysis puts it, a pause that feels harmless in a text interface makes a caller repeat themselves or assume the connection dropped. There is no loading spinner in a phone call. Perceived latency determines whether a voice agent feels competent or broken, and it is the failure mode users forgive least.
It is also not purely a model property. End-to-end latency in a voice agent is the sum of turn detection, network transport, tool execution and audio generation — so a fast model inside a slow pipeline still produces a slow experience. An agency building on this should be measuring p50 and p95 first-audio latency on its own stack, not reading "low latency" from a model card.
Google's design answer to this is the feature Extended Thinking is built around: instead of hiding the wait, the model talks through it — "Let me check that…" — and narrates progress. That is a genuinely good idea, and it is the most interesting UX contribution in this release. It converts dead air into evidence of work. The tradeoff is that Extended Thinking is slower by design, because it reasons aloud, and it costs roughly four times as much per hour of input audio.
What Reviewers Criticized
- The chess demo. Google's own demonstration showed the model losing to a basic checkmate pattern — widely called a bad look that undercuts the "most advanced" claim. When you demo your flagship reasoning model, the demo becomes the evidence.
- Benchmark provenance. Commentators questioned how much the evaluation harness contributed to the headline scores, particularly for figures run inside Google's own enterprise platform.
- No latency figures published for a product whose core promise is real-time conversation.
- Pricing opacity at launch. Google promised competitive pricing without publishing numbers, so "value" was unverifiable until third parties measured it.
- Base model regression on τ-Voice against the model it replaces.
- Fragmented rollout. Feature availability varies by surface and subscriber tier, with reports of users left on older model versions while seat prices rose.
- Watermarking questions. All audio output carries a SynthID watermark. Google's motive — provenance and abuse prevention — is reasonable, but at least one commentator questioned it, and it is worth knowing your generated audio is permanently marked.
- Admitted model-card caveats. Google's own documentation notes the models can hallucinate, that jailbreak protections are still maturing, and that they can occasionally slow down or time out. Forced tool use is not supported in all cases.
- Context and cost management is on you. Video frames are sent by default, so sessions can burn context and money quickly. Audio-plus-video sessions are capped at 2 minutes and audio-only at 15 minutes.
- Prior-generation hands-on was unflattering. The most substantial hands-on review available found the earlier Gemini Live answering questions it should have searched for, doubling down on hallucinations, and delivering "more filler than substance" — enough that the reviewer did not consider it worth the subscription premium. That is a warning about the product line rather than this release, but it is the closest thing to independent testing that exists right now.
The Honest Take
Google did something strategically sensible here and then undersold it badly. Splitting a voice line into a cheap volume tier and an expensive reasoning tier is the right product decision — voice agents have wildly different needs depending on whether they are answering "what is my account balance" or negotiating a multi-step refund, and pricing both at the same rate would have been worse for everyone.
The execution has two problems.
First, the cheap tier regressed. The base model scoring 30.1% on τ-Voice against its predecessor's 37.7% is not a rounding error — it is a 7.6-point decline on the benchmark that measures whether the model can complete tasks. Google is marketing that tier to developers as the cost-effective default, and the evidence says it should not be the default for anything agentic. If the reasoning tier is where the capability lives, then the cheap tier is a conversation model, not an agent model, and the marketing should say so.
Second, the transparency gap is real. No latency figures, no specific prices at launch, and headline benchmarks that are partly self-run. In voice, latency is the product. Publishing nothing about it while claiming "real-time" is the kind of omission that erodes trust in the numbers you did publish.
That said, two things here are genuinely good and worth crediting. The talk-through-the-wait design in Extended Thinking is a real UX insight — narrating background work converts silence from a failure signal into evidence of effort, and it addresses the single hardest problem in voice agents. And the ecosystem breadth — Agora, LiveKit, Pipecat, Vercel and others supporting the Live API from launch — materially shortens time-to-production for teams building on it.
What I would actually do: if you are building a voice agent that needs to complete tasks with tools, start with Extended Thinking and budget roughly $3.50 per hour of input audio, and measure your own p50/p95 first-audio latency before committing. If you are building conversational voice at volume where the model mostly talks and rarely acts, the base model's price is genuinely attractive — but verify the τ-Voice regression against your own task mix rather than assuming the newer model is the better one.
Pros & Cons
Pros
- Extended Thinking tops the Speech to Speech Quality Index at 82.6, ahead of GPT-Live-1 Astra and Grok Voice Think Fast 2.0 High
- Reasons while speaking — narrates background work instead of leaving dead air, a genuine UX insight for voice agents
- Base tier is the cheapest input audio on the panel at ~$0.84/hour; $0.005/min in and $0.018/min out
- Native audio-to-audio, not a stitched STT-LLM-TTS pipeline
- 97 languages with mid-conversation switching
- Runs tools and API calls in the background while the conversation continues
- Near-real-time visual grounding — the model can see what you show it
- Strong launch ecosystem — Agora, LiveKit, Pipecat, Vercel, LangChain and more supporting the Live API
- SynthID watermarking on all audio output aids provenance
- 128K context with a clear model ID for the reasoning tier
- Partners named — Salesforce, Genspark, Lumeris
Cons
- The base model regressed on τ-Voice — 30.1% versus the older Gemini 3.1 Flash Live's 37.7%
- Base model places fifth on the quality index, below both GPT-Live-1 configurations
- No latency figures published at all for a product sold on real-time conversation
- Extended Thinking costs ~4x the base tier per hour of input audio ($3.50 vs $0.84)
- Extended Thinking is slower by design because it reasons aloud
- Several benchmarks are vendor-reported or self-run, including inside Google's enterprise platform
- No independent head-to-head testing of the two models exists yet
- Chess demo failure undercut the "most advanced" positioning
- No prices published at launch — only a "highly competitive" promise
- Fragmented rollout across surfaces and subscriber tiers
- Proactive audio permanently on; affective dialog removed from the API
- Session caps — 2 minutes with video, 15 minutes audio-only
- Video frames sent by default, so context and cost need active management
- Preview status, not general availability
- Model-card caveats — hallucination, maturing jailbreak protections, occasional timeouts, no forced tool use in all cases
- Client state is more complex for Extended Thinking — must track
interaction_status, not justturnComplete
Who Should Use It (And Who Shouldn't)
| If you are… | Verdict |
|---|---|
| Building a voice agent that completes multi-step tasks with tools | Use Extended Thinking. Budget ~$3.50/hour input audio and measure your own latency |
| Building high-volume conversational voice where the model mostly talks | The base model's price is genuinely attractive — but test the τ-Voice regression against your own task mix first |
| Needing a voice agent for task completion on a budget | Look carefully at the older Gemini 3.1 Flash Live. It scores higher on τ-Voice and may be cheaper still |
| Wanting the best available voice quality regardless of cost | Extended Thinking. It is #1 on the independent quality index |
| Needing verified latency numbers before committing | Wait. Google published none; measure on your own stack or wait for independent benchmarks |
| Working in a customer-service deployment | Evaluate carefully. τ-Voice-banking is 35.1% and self-reported; the stakes in CX are high and the independent evidence is thin |
| Needing multilingual voice with visual context | Strong fit. 97 languages with mid-conversation switching plus near-real-time visual grounding |
| Requiring general availability for production | Not yet. This is a preview with staggered rollout |
| Needing to remove the audio watermark | You cannot. All output carries SynthID |
Frequently Asked Questions
What is the difference between Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking?
They are two separate models built for different jobs, not two quality tiers of one model. Gemini 3.8 Live is the fast, cheap option aimed at scale and cost efficiency: it handles visual input in near real time, auto-switches among 97 supported languages mid-conversation, and runs tools and API calls in the background while the conversation continues. Gemini 3.8 Live Extended Thinking is built for high-complexity multi-step tasks and reasons while speaking, using verbal cues such as "Let me check that" and narrating progress on background work. The base model signals completion with turnComplete and returns to idle, while Extended Thinking requires tracking an interaction_status because it can emit filler speech while still reasoning — a meaningfully more complex client state model.
Is Gemini 3.8 Live better than its predecessor?
The base model appears to have regressed on agentic voice tasks. DataCamp reports that Gemini 3.8 Live scores 30.1% on τ-Voice against 37.7% for the older Gemini 3.1 Flash Live, and places fifth on the Speech to Speech Quality Index at 76.0 — below both GPT-Live-1 configurations. It should not be the default choice for agentic or tool-chaining voice workloads. Extended Thinking is a different story: it captured the top spot on Artificial Analysis's Speech to Speech Quality Index at 82.6, ahead of GPT-Live-1 Astra at 81.5 and Grok Voice Think Fast 2.0 High at 81.3.
How much does Gemini 3.8 Live cost?
Through the Live API, audio costs $0.005 per minute of audio input and $0.018 per minute of audio output — roughly $1.38 per hour for two-way audio at those rates. Artificial Analysis separately measures Extended Thinking at about $3.50 per hour of input audio, against roughly $0.84 per hour for the base Gemini 3.8 Live, which is the cheapest entry on that panel. Tier token pricing on the Live tier runs $0.75 per million text tokens in and $4.50 out, $3.00 per million audio tokens in and $12.00 out, and $1.00 per million for image and video input. Note that Google's own announcement published no specific figures, promising only a highly competitive price point, so early value claims were speculative.
Does Gemini 3.8 Live work as a voice agent for customer service?
That is the intended use case, but with caveats. Google reports Extended Thinking at 68.6% on τ-Voice for agentic task completion and 35.1% on Sierra's τ-Voice-banking, and says the models push the Pareto frontier on ServiceNow's EVA-Bench for complex workflows — though those figures are vendor-reported and several were run inside Google's own enterprise platform. The base model is not suitable for tool-chaining workloads given its τ-Voice regression. Developers should also budget for context and cost management: video frames are sent by default, audio-plus-video sessions are capped at 2 minutes, and audio-only sessions at 15 minutes. Latency matters disproportionately in voice — a pause that feels harmless in text makes callers repeat themselves or assume a dropped connection.
Why did Google publish no latency figures?
Google has not said. What is clear is that no first-audio timing, p50 or p95 was published for either model, for a release whose entire pitch is real-time conversation. It is worth noting that end-to-end latency in a voice agent is not purely a model property — it is the sum of turn detection, network transport, tool execution and audio generation — so a single model-level figure would be of limited use anyway. The practical guidance is to measure p50 and p95 first-audio latency on your own stack rather than reading "low latency" from a model card.
Is the audio watermarked?
Yes. All audio output carries a SynthID watermark. Google's stated motive is provenance and abuse prevention, which is a reasonable position, and the model card covers safety. But you should know that generated audio is permanently marked — if your use case requires unwatermarked output, this model cannot provide it.
How do I access the models?
Both are available via the Gemini API, Google AI Studio and Vertex AI, with private preview in Gemini Enterprise. The model ID for the reasoning tier is gemini-3.8-live-extended-thinking. Consumer surfaces differ: Search Live gets the base model for everyone, while Gemini Live gets Extended Thinking for everyone, Workspace Docs for Google AI Pro and Ultra subscribers, and Gmail and Keep for all Google AI subscribers. Both models are in preview, not general availability.
Which platforms support the Gemini Live API?
Developer platforms building on the Live API include Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents. Google also names Salesforce, Genspark and Lumeris as partners. For voice agents — which are integration-heavy products spanning transport, turn detection and tool execution — having this many established platforms supporting the API at launch materially shortens the path to production.
Sources
- Google — Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking (September 15, 2026)
- 9to5Google — Gemini 3.8 Live Extended Thinking powers Gemini Live, Gmail & Keep
- DataCamp — Gemini 3.8 Live: features, benchmarks, pricing and access
- Agora — Gemini 3.8 Live vs Extended Thinking: Which Is Better for Voice AI?
- WaveSpeed — Gemini 3.8 Live vs Extended Thinking: Which Should You Use?
- OrcaRouter — Gemini 3.8 Live: Google Splits Its Voice Line in Two
- OrcaRouter — Gemini 3.8 Live vs Grok Voice Think Fast 2.0: The Gap
- BenchLM — Gemini 3.8 Live Extended Thinking: pricing, specs and sources
- TheBlockBeats — Gemini 3.8 Live released, tops voice leaderboard with 82.6 points
- FoneArena — Google rolls out Gemini 3.8 Live and 3.8 Live Extended Thinking
- InfoQ — 谷歌 Gemini 3.8 Live 要攻克语音 Agent 的沉默时刻
- Android Authority — I tried to love Gemini Live ahead of its Google Home debut
Disclosure: several benchmark figures cited here — τ-Voice, τ-Voice-banking, Big Bench Audio and EVA-Bench — are Google's own reporting, and at least one was run inside Google's enterprise platform. The Speech to Speech Quality Index from Artificial Analysis is an independent measure, but the base model's τ-Voice decline against Gemini 3.1 Flash Live comes from third-party analysis rather than from Google. No independent head-to-head testing of these two models against each other was available at the time of writing, and the most substantial hands-on review that exists covers the earlier Gemini Live rather than 3.8. Pricing was not published in Google's announcement and is drawn from third-party measurement. Both models were in preview status, with staggered rollout, so availability and pricing may change. Verify current figures on Google's official Gemini API pricing page and model card before building on this. This article reflects analysis of published information and does not represent hands-on testing by the author.
评论
发表评论