Google shipped two new voice AI models on September 15, 2026: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The announcement, posted to Google’s official blog under the byline of principal engineer Tom Ouyang and Gemini Audio Team staff member Malini Jaganathan, frames the pair as Google’s “most advanced live dialogue models yet.” That’s a bold claim in a crowded field, but the numbers Google published back it up in at least one respect: Gemini 3.8 Live Extended Thinking now holds the top overall spot on Artificial Analysis’ Speech to Speech Quality Index, a widely watched independent scoreboard for conversational AI.
The timing matters. This is the second Gemini 3.8 release inside two weeks – Google shipped Gemini 3.8 Flash and 3.8 Flash Cyber (see our coverage of the Gemini 3.8 Flash launch and the Gemini 3.8 Flash Cyber debut) on September 2, 2026 – and it lands squarely in the middle of an accelerating race among Google, OpenAI, and Anthropic to own the voice interface layer of consumer and enterprise AI. Search Engine Roundtable confirmed the rollout is already live in Google Search Live as of the announcement, meaning ordinary users are talking to the new model today, not in some future beta window.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Google Actually Announced on September 15
Google’s blog post describes two distinct models built for different jobs. Gemini 3.8 Live is positioned as the volume play: “built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding,” according to the announcement. Gemini 3.8 Live Extended Thinking is the heavier-duty sibling, described as “built for high-complexity tasks, with increased intelligence and multi-step reasoning.” As of September 2026, both have graduated out of preview naming: Google lists them in the Gemini API as the stable model versions gemini-3.8-live and gemini-3.8-live-extended-thinking, respectively, signaling that developers can build against them as production-ready identifiers rather than experimental snapshots. Google’s own API documentation goes a step further, now naming Gemini 3.8 Live as the platform’s default low-latency voice model as of the September 2026 documentation update – a designation that effectively makes it the baseline developers land on unless they explicitly opt into the heavier Extended Thinking variant.
Both models are aimed at what Google calls “near real-time reasoning” for voice agents – the ability to hold a natural spoken conversation while simultaneously running background logic, calling external tools, and pulling in visual context, without breaking conversational flow. That’s a meaningfully different design goal than a standard chatbot: Google’s framing throughout the post treats the microphone, not the keyboard, as the primary interface.
Gemini 3.8 Live handles visual inputs in near real time, according to the announcement, and automatically detects and switches between 97 supported languages mid-conversation – a feature aimed squarely at global enterprise deployments where a single voice agent might need to serve customers who switch between English, Spanish, and Hindi in the same call. It also executes tool calls and API requests in the background while the conversation continues, so the model can acknowledge a request verbally and keep talking while the underlying task finishes.
Inside Gemini 3.8 Live Extended Thinking
Extended Thinking is the more interesting engineering problem Google says it solved here. Reasoning models typically pause, think, and then respond – a workable pattern for text, but an awkward one for voice, where silence reads as a dropped connection or a confused assistant. Google’s answer is a model that “reasons and speaks simultaneously,” per the announcement, using conversational filler patterns like “Let me check that…” to acknowledge a request naturally while it works, followed by live progress narration as it walks through a multi-step background task. Google’s September 2026 technical documentation gives this behavior a formal name – interleaved reasoning – describing it as the mechanism that lets the model weave background “thinking” steps between spoken turns rather than blocking the conversation while it works.
Google demonstrated the pattern with several use cases in the launch post: an employee-onboarding assistant that answers live questions using visual context, a version of the model playing chess in near real time using visual reasoning, a workflow that turns raw sketches plus spoken feedback into functional React components, and a scenario where the model coordinates multi-step travel bookings and asynchronous function calls without interrupting the live conversation. None of these are lab curiosities – they’re the kind of workflows enterprise software vendors have spent two years trying to bolt onto older-generation voice assistants with limited success.
The Benchmark Numbers Behind the Claims
Google backed the announcement with a specific set of third-party and internal benchmark results rather than vague marketing language. The headline figure is Gemini 3.8 Live Extended Thinking’s score of 82.6 on Artificial Analysis’ Speech to Speech Quality Index, which the company says is enough to take the number one overall spot on that index. On agentic task completion, the model scored 68.6% on the τ-Voice benchmark and 35.1% on Sierra’s more demanding τ-Voice-banking benchmark, which tests whether a voice agent can complete realistic financial-services tasks end to end. On audio reasoning specifically, Google cited a 97.7% score on Big Bench Audio.
Gemini 3.8 Live, the lighter and cheaper of the two models, took second place in the Speech Agent Arena, a head-to-head user-preference ranking rather than a task-accuracy benchmark. Google also referenced results on ServiceNow’s EVA-Bench, a benchmark built specifically to evaluate voice agents, saying its models “push the Pareto Frontier for complex workflows by successfully balancing accuracy with conversational quality” – though the company did not publish a single numeric EVA-Bench score, noting only that the test was run on the Live API within the Gemini Enterprise Agent Platform.
| Benchmark | Model Tested | Result | What It Measures |
|---|---|---|---|
| Speech to Speech Quality Index (Artificial Analysis) | Gemini 3.8 Live Extended Thinking | 82.6 (#1 overall) | Overall voice conversation quality |
| τ-Voice | Gemini 3.8 Live Extended Thinking | 68.6% | Agentic task completion via voice |
| τ-Voice-banking (Sierra) | Gemini 3.8 Live Extended Thinking | 35.1% | Complex banking task completion via voice |
| Big Bench Audio | Gemini 3.8 Live Extended Thinking | 97.7% | Audio reasoning accuracy |
| Speech Agent Arena | Gemini 3.8 Live | 2nd place | User preference, head-to-head |
| EVA-Bench (ServiceNow) | Both models (Live API) | Pareto Frontier leader (no public score) | Accuracy vs. conversational quality balance |
Worth noting: these are Google-selected and, in the case of τ-Voice, τ-Voice-banking, and Big Bench Audio, Google-reported figures rather than fully independent audits. The Artificial Analysis and Speech Agent Arena rankings are run by outside organizations, which gives those two data points more weight as neutral confirmation. Independent replication of the τ-Voice and Big Bench Audio scores has not yet appeared as of publication.
Who Gets Access, and When
Google structured the rollout in three tiers – developers, enterprises, and general consumers – and both models started rolling out the same day as the announcement, not on some future date. That immediacy is itself notable: plenty of AI launches promise availability “in the coming weeks,” but Google confirmed live availability in Search Live starting September 15.
| Audience | Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking |
|---|---|---|
| Developers | Gemini API and Google AI Studio | Gemini API and Google AI Studio |
| Enterprises | Private preview in Gemini Enterprise; coming soon to Gemini Enterprise for Customer Experience | Private preview in Gemini Enterprise; coming soon to Gemini Enterprise for Customer Experience and Workspace business customers |
| Consumers | Search Live | Gemini Live app; Docs for Google AI Pro/Ultra subscribers; Gmail and Keep for all Google AI subscribers |
The consumer split is deliberate. Gemini 3.8 Live, the cheaper and faster model, goes into Search Live, where Google needs to serve enormous query volume at low latency and lower cost. Extended Thinking, the more expensive and capable model, goes into Gemini Live and Workspace apps where users are more likely to be working through genuinely complex tasks – drafting a document by voice in Docs, triaging a cluttered inbox in Gmail, or organizing notes in Keep – and where Google can gate access by subscription tier (Google AI Pro and Ultra for Workspace, versus all Google AI subscribers for Gmail and Keep).
The Third and Fourth Gemini 3.8 Release in Two Weeks
Context matters here. Google announced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026 – the workhorse and cybersecurity-focused members of the 3.8 family – meaning the Live and Live Extended Thinking models arrived just 13 days later as the third and fourth named releases under the Gemini 3.8 banner. Gigazine, the Japanese tech outlet that covered the September 2 launch, flagged that Gemini 3.8 Flash pricing is expected to roughly double starting in 2027, and Google has since put a hard date on that window: introductory pricing for both Gemini 3.8 Flash and Gemini 3.8 Flash Cyber is set to expire on December 31, 2026, a detail that suggests Google is using today’s aggressive rollout cadence partly to lock in developers on current pricing before the increase hits at the start of next year.
| Date | Release | Focus |
|---|---|---|
| September 2, 2026 | Gemini 3.8 Flash | General-purpose workhorse model for agentic tasks and software engineering |
| September 2, 2026 | Gemini 3.8 Flash Cyber | Cybersecurity-specialized variant |
| September 15, 2026 | Gemini 3.8 Live | Cost-efficient live dialogue with visual grounding |
| September 15, 2026 | Gemini 3.8 Live Extended Thinking | High-complexity multi-step reasoning for voice agents |
Digital Applied’s AI model tracker, which logs dated model releases across the industry, described the September 2 Flash launch as Google’s “third Flash release in six weeks by Google’s count,” pointing to a broader pattern: Google is not shipping one flagship model a year anymore. It’s iterating in public on a near-weekly cadence, treating model releases the way a SaaS company treats feature flags – ship fast, gather usage data, revise. That cadence has roots further back than September: Google first announced Gemini Live’s inline chat integration at I/O on May 20, 2026, four months before the 3.8 family arrived, meaning the September releases are less a sudden pivot than the latest step in a live-voice roadmap Google has been building publicly since the spring.
Pricing and API Access for Developers
Google’s announcement did not publish standalone per-token or per-minute pricing for Gemini 3.8 Live or Extended Thinking in the launch post itself, but the Gemini API pricing page has since caught up: as of September 2026, Google bills Gemini 3.8 Live audio input at $0.005 per minute and audio output at $0.018 per minute. What it did confirm at launch is availability through the Gemini API and Google AI Studio starting immediately, alongside the Gemini API release notes, which list the new model identifiers in the official changelog. TestingCatalog separately reported that the Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking model names had already surfaced on the Google Cloud Platform console’s quota and metrics page hours before the public announcement went live – a common tell that a launch is imminent, since backend billing and quota infrastructure typically ships slightly ahead of the public-facing blog post.
For enterprise buyers evaluating cost, the closest available reference point is Gemini 3.8 Flash, which launched September 2 at an introductory per-token rate matching its predecessor, Gemini 3.7 Flash, according to Google’s own post – though that introductory rate is confirmed to run only through December 31, 2026, after which Google’s own guidance points to pricing roughly doubling. The premium for real-time audio is now visible in the published numbers: with Gemini 3.8 Live billed at $0.005 per minute for audio input and $0.018 per minute for audio output as of September 2026, and both it and Extended Thinking listed under the stable gemini-3.8-live and gemini-3.8-live-extended-thinking identifiers, enterprise buyers finally have a concrete basis for cost modeling rather than the introductory-pricing guesswork available at launch – and a hard deadline before the current rates expire.
How This Compares to GPT-6 Astra and Claude Opus 5
Google’s timing puts real pressure on OpenAI and Anthropic, both of which have their own frontier models – GPT-6 Astra and Claude Opus 5, respectively – competing for enterprise AI budgets. Neither OpenAI nor Anthropic has published a benchmark result directly comparable to Artificial Analysis’ Speech to Speech Quality Index for their current flagship models, so a precise apples-to-apples score comparison isn’t possible from public data today. What is clear is the strategic gap Google is trying to open: by shipping a dedicated, benchmark-topping voice-first model family rather than treating voice as a bolt-on feature of a general-purpose text model, Google is betting that live conversational AI is different enough from text-based reasoning to deserve its own architecture and its own release cadence.
That bet has a historical precedent worth remembering. Google’s Search Live product itself only exists because Google concluded that bolting a chat window onto Search wasn’t enough – it needed a model built to sustain a spoken back-and-forth. Reports on the broader voice AI market suggest OpenAI’s real-time voice stack and Amazon’s Alexa+ remain the two most direct competitors for developers building production voice agents, a rivalry we tracked previously in our comparison of ChatGPT Voice, Grok Voice, and Gemini Live pricing, though performance comparisons across all three platforms change frequently enough that any specific numeric comparison risks being outdated within weeks.
Why Google Is Racing to Own the Voice Interface
The market logic here is straightforward once you separate it from the benchmark noise. Every major AI lab has now shipped, or is actively building, a “computer use” or agentic product that can take actions on a user’s behalf rather than just answering questions. Voice is the interface layer that makes agentic AI feel natural rather than mechanical – nobody wants to type a full paragraph of instructions to book a flight or troubleshoot a printer. Google’s decision to route Extended Thinking into Docs, Gmail, and Keep specifically (rather than only the standalone Gemini app) signals an intent to make voice-driven task execution the default interaction mode across Workspace, not a novelty feature buried in settings.
There’s also a defensive angle. Search Live getting Gemini 3.8 Live specifically – not the flagship Extended Thinking model – suggests Google is optimizing its highest-traffic consumer surface for cost and latency first, reserving the more expensive, more capable model for lower-volume, higher-value enterprise and prosumer contexts. That’s a rational allocation of compute given how many queries Search processes daily, but it also means the “most advanced” model Google is touting in its own headline isn’t the one most people will actually talk to.
The Developer Ecosystem Building on Gemini Live
Google didn’t launch this in isolation. The announcement names a specific set of developer platforms – Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents – that already integrate with the Gemini Live API, handling the real-time media streaming infrastructure so individual developers don’t have to build WebRTC pipelines from scratch. That’s a meaningful signal: infrastructure providers don’t invest engineering time supporting a new model API unless they expect real developer demand to follow, and having seven named integration partners at launch suggests Google pre-briefed this ecosystem well before September 15.
On the enterprise side, Google named Salesforce, Genspark, and Lumeris as partners it says are already excited about the new models’ latency, fluidity, and tool-calling capabilities. Lumeris in particular is a healthcare-focused technology company, which hints at where Google expects early enterprise traction: high-stakes verticals like healthcare and financial services where a voice agent needs to reliably execute multi-step tasks (scheduling, eligibility checks, claims status) without losing the thread of a live conversation.
Safety and the SynthID Watermark
Every piece of audio these models generate carries a SynthID watermark, according to Google’s announcement. The company describes it as an imperceptible signal woven directly into the audio output, designed to keep AI-generated audio detectable as a defense against misinformation and impersonation. That’s a meaningful commitment given how easily voice-cloning and synthetic-audio scams have already caused real financial harm across the industry – a watermarking layer at the model level, rather than as an optional add-on, raises the baseline for detectability across every surface where these models ship, from Search Live to enterprise call centers.
Google pointed to its model card for further detail on the safety and responsibility approach behind both models, though it did not disclose specific red-teaming results, jailbreak-resistance figures, or independent third-party safety audits in the launch post itself.
Market Impact: What This Means for Enterprise AI Budgets
For enterprise buyers already juggling AI model contracts, this launch adds another decision point to an already crowded field. Companies that adopted Gemini 3.8 Flash for text-based agentic workflows in early September now have a same-family option for voice, which simplifies vendor management if they choose to stay within Google’s ecosystem rather than mixing providers for text versus voice. The private-preview gating on Gemini Enterprise access means large customers will get first access, while smaller developers building through the public Gemini API can start integrating immediately without a sales conversation.
The bigger structural shift is what this does to the build-versus-buy calculus for voice AI. Before dedicated live-dialogue models with published, independently verified benchmark rankings existed, most enterprises building voice agents had to stitch together separate speech-to-text, reasoning, and text-to-speech pipelines – adding latency and cost at every seam. A single model handling perception, reasoning, and generation natively, as Gemini 3.8 Live claims to do, removes at least two of those seams, which should meaningfully lower both latency and integration cost for teams building production voice agents going forward.
Predictions: Where Live Voice AI Goes From Here
Based on the pattern Google has established with the Gemini 3.8 family’s release cadence and the competitive pressure this launch puts on rivals, a few developments look likely over the next two to three quarters:
- Expect a response launch from OpenAI or Anthropic within weeks, not months. Google’s pattern of shipping model updates every few weeks has already compressed the industry’s typical response window, and a benchmark-topping voice model is the kind of headline competitors rarely let sit unanswered for long.
- Gemini Enterprise’s private preview for Live and Extended Thinking will likely open to general availability before the end of 2026, consistent with how quickly Google moved Gemini 3.8 Flash from announcement to broad availability.
- Pricing transparency will improve once developer adoption ramps up – Google typically publishes granular per-token or per-minute pricing pages once a model exits its initial rollout phase, as it did with prior Gemini Flash releases.
- Voice-first agentic workflows will become a standard RFP requirement for enterprise software vendors in healthcare and financial services, the two verticals Google’s own named partners (Lumeris, and implicitly Salesforce’s customer-service use cases) suggest are seeing the earliest traction.
- Independent benchmark scrutiny of the τ-Voice and Big Bench Audio figures will likely surface within the next month, as third-party evaluators typically attempt to replicate vendor-reported scores once a model reaches public API access.
These are analytical projections based on Google’s disclosed rollout pattern and competitive dynamics, not confirmed roadmap items from Google, OpenAI, or Anthropic.
The Bigger Picture: Voice as the Next AI Platform War
Zoom out and this launch reads less like a single product announcement and more like a marker in an ongoing platform war. Text-based chatbots reached a rough performance plateau over the past year – differences between frontier text models increasingly come down to price and context-window size rather than dramatic capability jumps. Voice is the frontier where meaningful differentiation is still possible, because the technical problems (near real-time latency, natural turn-taking, background task execution without breaking conversational flow) remain genuinely hard.
Google topping an independent voice-quality index gives it a concrete talking point in that fight. Whether Gemini 3.8 Live and Extended Thinking translate that benchmark lead into actual usage share against OpenAI’s and Anthropic’s voice offerings will depend on factors the launch post doesn’t address: real-world latency under load, pricing once introductory periods end, and how reliably these models perform outside Google’s own curated demo scenarios of onboarding assistants, chess games, and sketch-to-code workflows.
Frequently Asked Questions
What is Gemini 3.8 Live?
Gemini 3.8 Live is a voice AI model Google announced on September 15, 2026, described as built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding, according to Google’s official blog post.
What is Gemini 3.8 Live Extended Thinking?
It’s the higher-capability companion model, built for high-complexity tasks with increased intelligence and multi-step reasoning. It can reason and speak simultaneously, using natural verbal cues to acknowledge requests while working through background tasks.
How do I access Gemini 3.8 Live as a developer?
Both models are available immediately through the Gemini API and Google AI Studio, according to Google’s announcement. Model identifiers also appear in the official Gemini API changelog.
Is Gemini 3.8 Live available to regular consumers?
Yes. Gemini 3.8 Live is live now in Google Search Live. Gemini 3.8 Live Extended Thinking is available in the Gemini Live app, and for Google AI Pro and Ultra subscribers in Workspace’s Docs, and for all Google AI subscribers in Gmail and Keep.
How many languages does Gemini 3.8 Live support?
Google says the model automatically detects and transitions between 97 supported languages mid-conversation.
How does Gemini 3.8 Live compare to GPT-6 Astra or Claude Opus 5 on voice?
Neither OpenAI nor Anthropic has published a directly comparable benchmark score for their flagship models on independent voice-quality indices like Artificial Analysis’ Speech to Speech Quality Index, so a precise numeric comparison isn’t currently possible from public data.
Is AI-generated audio from these models watermarked?
Yes. Google says all audio generated by its AI products, including Gemini 3.8 Live and Extended Thinking, carries an imperceptible SynthID watermark intended to keep AI-generated content detectable.
Which developer platforms already support the Gemini Live API?
Google named Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents as platforms enabling developers to build voice-driven interfaces on top of the Gemini Live API at launch.
Related Coverage
- Siri Runs 5 Gemini-Based Models, Not Google's App [2026]
- Googlebook Pre-Orders Open Sept. 21: $999 Gemini Laptops
- OpenAI Realtime API: gpt-live-1 Setup in 12 Steps [2026]
- Google-Blackstone's Crux AI Poaches Meta's Duong [2026]
- Gemini Lands on Windows With 1 Shortcut, No OS Hooks [2026]
