Gemini 3.8 Live: Google's real-time voice at $1.38/hour — background reasoning and tool execution without interrupting the conversation
🔎 AI voice stops going silent to think
On September 15, 2026, Google DeepMind launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two native speech-to-speech voice models purpose-built for voice agents (Unite.ai). Behind the official headline lies a fundamental problem that AI voice has been dragging around for two years: an agent had to choose between answering fast and thinking accurately. The moment it had to query a CRM or a database, it went silent. Users hung up.
Gemini 3.8 Live tackles this bottleneck head-on: background reasoning, tool execution, and API calls without interrupting the conversation, with a model that narrates its progress out loud (Google blog). The Alchemy Lab Horizon September 16 briefing immediately ranks it among the week's major launches.
The second front, less glamorous and more decisive: pricing. Roughly $1.38 per hour of conversation versus at least $3 at OpenAI (AI Daily Post). At that level, real-time voice stops being a demo and becomes a line item in your costs. What this changes for your phone agents, the math against a human call-center agent, and the state of the real-time war: that's what this article is about.
The Essentials
- Two models, not one: Gemini 3.8 Live (cost efficiency) and Gemini 3.8 Live Extended Thinking (near-real-time reasoning), launched on September 15, 2026.
- #1 on Artificial Analysis' Speech to Speech Quality Index: 82.6, ahead of OpenAI's GPT-Live-1 Astra (81.5) and xAI's Grok Voice Think Fast 2.0 (81.3).
- Aggressive pricing: $0.005/min for input and $0.018/min for output, roughly $1.38/h — and 40% less than OpenAI on Extended Thinking ($3.50 vs. $5.83/h).
- Latency under 200 ms versus 450–1,000 ms for traditional cascaded ASR + LLM + TTS architectures.
- The end of "dead air": asynchronous tool execution, progress narration, mid-conversation transitions between 97 languages.
- An acknowledged limitation: only 35.1% on τ-Voice-banking — complex calls remain human territory.
Recommended tools
The current building blocks for building a voice agent, with prices recorded on September 16, 2026 (check the official websites before making any commitment):
| Tool | Main use | Price (Sept. 2026) | Ideal for |
|---|---|---|---|
| Gemini 3.8 Live | High-volume real-time voice | $0.005/min input, $0.018/min output (~$1.38/h) | Tier 1 support, appointments, hotlines |
| Gemini 3.8 Live Extended Thinking | Voice + reasoning + tools | $3.50/h | Multi-step tasks, complex workflows |
| GPT-Live-1 Astra (OpenAI) | Premium full-duplex voice | $5.83/h | OpenAI ecosystem already in place |
| Grok Voice Think Fast 2.0 (xAI) | Low-latency voice | $4.80/h | Short conversations, responsiveness |
| PhoneLLM Alpha-1 | Open-weights voice, self-hosted | $0.0025/min self-hosted | Cost control, internal data |
| VoiceStudio | 100% local speech synthesis | Free, open source | Privacy, on-premise deployment |
Two models, one architecture: what Google actually launched
Google isn't releasing "one" voice model, but two variants of the same native speech-to-speech family: Gemini 3.8 Live, optimized for cost efficiency at scale, and Gemini 3.8 Live Extended Thinking, built for reasoning. Both are available as of September 15, 2026 on Gemini API, AI Studio, Gemini Enterprise, Search Live, Gemini Live, and Workspace (Unite.ai).
The announcement comes from Tom Ouyang, principal engineer, and Malini Jaganathan, of the Gemini Audio Team. The detail that matters for developers: a native speech-to-speech architecture that replaces the long-standing ASR + LLM + TTS cascade (ByteIota).
Why is this a structural shift? In a cascade, the voice is transcribed into text, the LLM reasons over this impoverished text, and then a synthesis engine regenerates speech. Three steps, three sources of errors, and a cumulative latency of 450 to 1,000 ms. Speech-to-speech processes the audio end to end: latency drops below 200 ms, the threshold where a conversation stops feeling artificial (ByteIota).
Two Gemini 3.8 releases in two weeks
Gemini 3.8 Flash and Flash Cyber on September 2, Live on the 15th (Tech-Insider). This pace is a message as much as a roadmap: Google is flooring the accelerator on real time, a territory OpenAI believed it held with its full-duplex voice.
"Near real-time reasoning": the end of dead air?
Yes — and it is, by far, the most important innovation of this launch. Far more than the benchmarks. Extended Thinking reasons in the background, executes tools, and calls APIs without interrupting the conversation (Google blog).
The classic dead air scenario: "check if my plan includes international calling". The agent goes silent for eight seconds while the CRM call runs. The user repeats themselves, sighs, hangs up. That silence was the number one killer of voice agents in production.
Google's answer comes down to two complementary mechanisms:
- Non-blocking asynchronous tool execution: the conversation continues while the API call runs (ByteIota).
- Progress narration: the model announces out loud what it's doing — "I'm pulling up your file, two seconds" — during multi-step tasks, without breaking the flow (Unite.ai).
My take: the narration is a UX tour de force more than a model one. It doesn't reduce actual latency, it makes it bearable. This is exactly what no ASR + LLM + TTS cascade can do properly — you'd have to script every case by hand, and nobody does.
Benchmarks: Google takes the lead, with asterisks
Gemini 3.8 Live Extended Thinking snags first place in Artificial Analysis' Speech to Speech Quality Index with 82.6, ahead of OpenAI's GPT-Live-1 Astra (81.5) and xAI's Grok Voice Think Fast 2.0 (81.3) (Unite.ai).
| Benchmark | Gemini 3.8 Live Extended Thinking | GPT-Live-1 Astra (OpenAI) | Grok Voice Think Fast 2.0 (xAI) |
|---|---|---|---|
| Speech to Speech Quality Index (Artificial Analysis) | 82.6 — #1 | 81.5 | 81.3 |
| τ-Voice (Sierra) | 68.6% | n/a | n/a |
| τ-Voice-banking (Sierra) | 35.1% | n/a | n/a |
| Big Bench Audio | 97.7% | n/a | n/a |
| Speech Agent Arena (user preference) | 2nd | n/a | n/a |
| Hourly price (Sept. 2026) | $3.50 | $5.83 | $4.80 |
n/a: not disclosed in public sources as of September 16, 2026.
Three caveats before concluding too hastily that this is a done deal:
- User preference doesn't follow the score. On the Speech Agent Arena, the user preference leaderboard, Gemini is only second (AI Chat Daily). The highest-rated model isn't always the one humans prefer when listening.
- ServiceNow's EVA-Bench is a self-evaluation. The "best Pareto frontier of accuracy/conversational quality" claimed by Google is measured via its own platform, the Gemini Enterprise Agent Platform. Interesting, but take it with the usual grain of salt (AI Chat Daily).
- 35.1% in a banking environment is a ceiling, not a springboard. On τ-Voice-banking's complex scenarios, the world's #1 still fails two times out of three. For simple support, it's ready. For autonomous advice, it's not.
The price: $1.38/hour, an unapologetic undercutting of OpenAI
$0.005/minute for audio input, $0.018/minute for output: an hour of conversation costs about $1.38 at Google, versus at least $3 at OpenAI with GPT-Live-1 (AI Daily Post, September 2026).
| Model | Audio input | Audio output | Estimated hourly cost (Sept. 2026) |
|---|---|---|---|
| Gemini 3.8 Live (Google) | $0.005/min | $0.018/min | ~$1.38 |
| Gemini 3.8 Live Extended Thinking | n/a | n/a | $3.50 |
| GPT-Live-1 Astra (OpenAI) | $0.05/min | n/a | $5.83 |
| Grok Voice Think Fast 2.0 (xAI) | n/a | n/a | $4.80 |
The number that hurts OpenAI: audio input. $0.30/hour at Google versus $3.00/hour at OpenAI — a 10x factor on the heaviest line item of the bill, since a voice agent listens far more than it speaks. According to AI Daily Post, this gap turns voice agents "from experimental demos into viable products at scale". The September 16 digest from AI0.news points to the same dynamic: the competition has shifted from benchmarks to the bill.
In real-world billing, ByteIota prices out a support bot handling 10,000 hours/month: $58,300 with GPT-Live-1, $35,000 with Extended Thinking (−40%), $8,400 with the base model. That works out to an effective cost of $0.84/hour at high volume — below the listed price, since real conversations contain silences.
An example at SMB scale (our own calculation based on these price grids): 5,000 calls of 6 minutes per month, i.e. 500 hours — roughly $690 on the base model, $1,750 on Extended Thinking, $2,915 on GPT-Live-1. The choice of model is, in itself, a budget decision.
One methodological caveat: the "$1.38/hour" figure is a press estimate based on a typical call profile, whereas the official price grid is displayed by the minute and distinguishes input from output. Always recalculate using your own call recordings.
Phone agents: what actually changes
Three measurable changes for call centers: latency finally below the naturalness threshold, the end of silences during requests, and an hourly cost that drops below one euro at volume. Here's what it looks like in practice.
Latency, the foundation of naturalness. From 450–1,000 ms with cascading setups to under 200 ms natively (ByteIota). Below this threshold, the exchange no longer sounds "robotic." It's invisible on a slide, obvious on the very first customer call.
Tools no longer interrupt the conversation. Order verification, availability checks, appointment booking: everything runs in the background while the agent keeps talking and narrates its progress. The gain will be measured in call abandonment rates, not in benchmarks.
Multilingual without transfers. Automatic transitions between 97 languages mid-conversation (Unite.ai). A call that starts in French and switches to English without putting the caller on hold: a killer argument for international support.
Distribution everywhere. Search Live, Gemini Live, Workspace, Gemini Enterprise, AI Studio, API (Unite.ai): Google plugs real-time voice into everywhere businesses already work. Less spectacular than a leaderboard, more decisive for adoption.
On the flip side: τ-Voice at 68.6% means a third of complex voice tasks still fail. Route the simple cases to the agent, keep human escalation for the rest. And if you're assembling the pipeline yourself, our selection of free AI APIs is the simplest entry point for prototyping — complemented by the best free LLMs for the text side.
$1.38/hour vs. a call center agent: the real math
Even with Extended Thinking at $3.50/hour, a voice agent costs 5 to 6 times less than a human call center agent. But the 35.1% on τ-Voice-banking is a reminder that full replacement isn't happening any time soon.
The raw math: in France, a call center agent costs at least €15 to €20 per hour fully loaded — salaries, payroll taxes, facilities, management (ballpark figure). Against Gemini's $0.84 to $3.50/hour, the gap ranges from 1 to 5 at the very least, up to 1 to 20 at volume on the base model.
But the real question isn't "human or AI." It's: which hours? A support center concentrates most of its volume on recurring tasks — order tracking, password resets, appointment booking. There, Gemini 3.8 Live pays for itself immediately. Sensitive calls — complaints, financial advice, customers in distress — stay human, because the model still fails two times out of three in banking environments and isn't even first in user preference.
My take: executives who present this as "we're eliminating the floor" will be buying problems for their customers and disputes for their legal departments. The honest reading of the Sierra and Artificial Analysis numbers: AI takes the volume, humans keep the complexity and the control. Less sexy in a press release, more solid on a balance sheet.
The Real-Time War: OpenAI, xAI and the open weights counterattack
Google is attacking OpenAI head-on on its own turf — full-duplex voice — with a higher score and a price 40% lower. Tech-Insider places this launch at the heart of a three-way race between Google, OpenAI and Anthropic in real-time voice. Here's where the players stand:
OpenAI holds the quality, not the price. GPT-Live-1 Astra: 81.5 on the quality index, $5.83/h (ByteIota). OpenAI opened fire with GPT-Live's full-duplex voice, which finally allowed agents to listen and speak simultaneously (our analysis), then with GPT-Realtime-2 and its three voice models capable of reasoning, translating and transcribing in real time (our coverage). With two Gemini 3.8 releases in two weeks, Google is making it clear that no one will get a six-month head start.
xAI sits in mid-table. Grok Voice Think Fast 2.0: 81.3 and $4.80/h (ByteIota). Neither the best score nor the best price — an uncomfortable position to be in.
Open weights are lowering the floor. PhoneLLM Alpha-1 matches closed models at $0.0025 per minute when self-hosted (our article) — roughly $0.15/h, ten times less than base Gemini 3.8 Live. And on speech synthesis, VoiceStudio, the 100% local alternative to ElevenLabs, has surpassed 12,900 stars (our feature).
To keep control of your data and your costs, our comparison of the best LLMs to run locally rounds out these voice building blocks. And this battle plays out month by month: you can follow it in our monthly comparison of the best LLMs.
❌ Common Mistakes
Mistake 1: Comparing prices in different units
"$0.05/minute" and "$1.38/hour" cannot be compared as-is: the first is an API rate card, the second a press estimate. The fix: extract your average listening and speaking minutes from your recordings, apply the official input/output rate card, then compare on an identical hourly basis.
Mistake 2: Routing everything to Extended Thinking
The benchmark leader costs $3.50/h; the base model, roughly a quarter of that in volume. For a hotline with a high density of simple questions, that's a waste. The fix: two routes — recurring volume goes to Gemini 3.8 Live, multi-step journeys with tool calls go to Extended Thinking.
Mistake 3: Believing that background reasoning makes the model instant
Narration masks latency, it doesn't eliminate it: a multi-step verification still takes a few seconds. The fix: explicitly write progress scripts ("I'm checking, give me two seconds"), test for API failures, and plan for conversation resumption when the tool goes down.
Mistake 4: Deploying autonomously on complex journeys
35.1% success rate on τ-Voice-banking: two failures out of three in complex banking scenarios. Launching an unsupervised agent on top of that is buying yourself disputes. The fix: a systematic human escalation threshold, call logs retained, GDPR-compliant information and consent — compliance is part of the business case.
❓ Frequently Asked Questions
Where is Gemini 3.8 Live available?
On Gemini API, AI Studio, Gemini Enterprise, Search Live, Gemini Live, and Workspace, according to the September 15, 2026 announcement (Unite.ai). Developers test via AI Studio, enterprises deploy via Gemini Enterprise on their internal data — and Search Live or Gemini Live also expose it to the general public.
What's the difference between Gemini 3.8 Live and Extended Thinking?
The base model is positioned for cost efficiency at scale: roughly $1.38/h estimated, $0.84/h effective at volume. Extended Thinking adds background reasoning, asynchronous tool execution, and narration, for $3.50/h. It's this version that tops Artificial Analysis' quality index (82.6).
How much does a voice agent cost at 10,000 hours per month?
According to ByteIota (September 2026): $8,400 with base Gemini 3.8 Live, $35,000 with Extended Thinking, $58,300 with OpenAI's GPT-Live-1. Switching from OpenAI to Google therefore saves up to nearly $50,000 per month — before even optimizing routing between the two Gemini models.
Does Gemini 3.8 Live handle multiple languages in a single call?
Yes: the model automatically switches between 97 languages within a single conversation, with no hold times or handoffs to another model. A call that starts in French and continues in English stays in the same flow — a strong point for international support.
Should you ditch GPT-Live-1 for Gemini 3.8?
On paper, the trade-off is clear: a higher score (82.6 vs 81.5) and a price 40% lower in Extended Thinking. But Gemini is only second in the Speech Agent Arena. Before migrating, test on your own calls: user preference can contradict the benchmarks.
Are there open source or local alternatives?
Yes, on two fronts: PhoneLLM Alpha-1, a self-hostable open-weights voice model at $0.0025 per minute, and VoiceStudio for 100% local speech synthesis, which has surpassed 12,900 stars. The quality holds up too — the price floor simply isn't what it was before this launch.
✅ Conclusion
Gemini 3.8 Live takes real-time voice from impressive-demo status to that of a calculable cost line: number one in speech-to-speech, reasoning without silences, $0.84 to $1.38/h — the bottleneck is no longer technical, it's organizational. To follow the Google/OpenAI/xAI match month by month, head over to our comparison of the best LLMs; and if your goal is to make these new use cases profitable, our hands-on review of 7 AI tools that earned me €300/month without coding is a good starting point.