Google just cut the price of a realtime voice conversation by more than half. Gemini 3.8 Live, the newest entry in Google’s voice-model line, prices audio input at $0.005 per minute and output at $0.018 per minute — a structure that lands an hour of back-and-forth conversation at roughly $1.38. GPT-Live-1, OpenAI’s answer and the successor to GPT-Realtime-2, charges a flat $0.05 per minute. Run the same hour through it and you’re at $3.00 or more.

That’s the headline, and it’s real. It’s also not the whole story. On Artificial Analysis’ Speech-to-Speech Quality Index, Gemini’s premium Extended Thinking tier edges out GPT-Live-1’s mid-tier configuration — barely. Standard tier against standard tier, the race flips, and GPT-Live-1’s base configuration wins the composite score outright. Add in the architecture gap — GPT-Live-1 runs full-duplex, Gemini 3.8 Live’s standard tier is turn-based — and “cheaper” and “better” stop being the same question. They’re not even measuring the same thing.

Quick Verdict

Aspect	Gemini 3.8 Live	GPT-Live-1
Best For	High-volume, cost-sensitive voice apps; multimodal sessions	Natural, interruptible conversation; premium support and sales lines
Standard Pricing	$0.005/min input, $0.018/min output	$0.05/min flat
Cost per hour (conversation)	~$1.38	~$3.00+
Premium tier ($/hour)	Extended Thinking: $3.50	Sol: $4.47
Speech-to-Speech Quality Index	82.6% (Extended Thinking)	81.5% (Astra, Medium)
Architecture	Turn-based (standard tier)	Full-duplex
Languages	97+	Fewer at launch
Extras	Background API calls, visual input while talking	Native interruption handling

Bottom line: Gemini 3.8 Live is the cheaper voice API by a wide margin and it wins one narrow benchmark at the top tier. GPT-Live-1 is still the better conversation partner for most standard-tier use cases, because full-duplex audio beats turn-based audio in the moments that actually annoy users — interruptions, overlapping speech, the parts of a real conversation nobody scripts.

---

What Actually Shipped

Google’s pitch for Gemini 3.8 Live is volume economics. Per-minute pricing that splits input and output — $0.005 and $0.018 respectively — rewards exactly the workload every contact center and voice-app team is trying to build: long sessions, lots of listening, shorter bursts of talking back. A support call that’s mostly the customer explaining a problem costs less than one where the model does most of the talking. That’s a deliberate pricing shape, not an accident.

OpenAI’s GPT-Live-1 keeps it simple instead: one flat rate, $0.05 per minute, no separate meter for direction. Easier to budget. Also, on a typical two-way conversation, meaningfully more expensive — about 2.2x the Gemini rate on the same hour, based on the published per-minute figures.

Both companies also shipped a second, premium tier aimed at harder reasoning during a live call — tool use, multi-step lookups, the kind of thing that used to force a handoff to a human agent. Here the pricing structure converges: both charge a flat hourly rate instead of splitting input and output.

Tier	Rate
Gemini 3.8 Live Extended Thinking	$3.50/hour
GPT-Live-1 Sol	$4.47/hour
Grok Voice Think Fast 2.0 High	$4.80/hour

Gemini undercuts both competitors here too. Against GPT-Live-1 Sol — built on the same Sol-line reasoning behind GPT-5.6 Sol — Extended Thinking runs about 22% cheaper per hour. Against Grok Voice Think Fast 2.0 High, it’s about 27% cheaper. Route meaningful volume through a reasoning-heavy tier and that gap compounds fast.

Why Is Gemini 3.8 Live So Much Cheaper Than GPT-Live-1?

1. Split metering rewards listening-heavy calls. Gemini bills input and output separately, and input is priced at roughly a quarter of output. Most real conversations lean toward listening, so the blended rate comes in lower than a flat per-minute charge would.
2. Google is subsidizing volume, not margin. The same land-grab logic behind Google AI Ultra’s price cut and the Gemini Spark rollout applies here — Google has been willing to run voice and agent products at thinner margins to win developer default status before the category settles.
3. OpenAI’s flat rate trades simplicity for headroom. A single per-minute number is easier to forecast and easier to sell into an enterprise contract, but it doesn’t reward the usage patterns — long listening stretches, short replies — that make voice agents cheap to run.
4. The reasoning tiers price differently than the standard tiers. Extended Thinking, Sol, and Think Fast 2.0 High all use flat hourly rates instead of split metering, which narrows the gap between vendors at the premium end even as it stays wide at the standard end.

Where Gemini 3.8 Live Wins

Price, By a Wide Margin

The math isn’t close. An hour of standard-tier conversation on Gemini 3.8 Live costs about $1.38. The same hour on GPT-Live-1 costs $3.00 or more. For a voice app running thousands of hours a month — appointment scheduling, order status lines, first-tier support triage — that difference is the line item that decides whether the product is profitable at scale, not a rounding error on an invoice.

The Extended Thinking Benchmark

On Artificial Analysis’ Speech-to-Speech Quality Index — a composite that scores naturalness, latency-adjusted comprehension, and task completion in a live voice session — Gemini 3.8 Live’s Extended Thinking tier posts 82.6%, narrowly ahead of GPT-Live-1 Astra at the Medium setting, which scores 81.5%. A one-point gap isn’t a rout. It is, at minimum, proof that Google’s reasoning tier isn’t just cheaper — it’s competitive at the top of the market, not just the bottom.

Multimodal While Talking

Gemini 3.8 Live can accept visual input mid-conversation — point a camera at something while you’re still talking and the model responds to what it sees without you having to stop, upload, and resume. It also supports background API calls, meaning the model can trigger a tool call or a lookup without breaking the audio stream to do it. And it covers 97+ languages, a wider net than GPT-Live-1 has at launch. None of that shows up in the per-minute price. All of it matters if your product needs to see as well as hear.

Where GPT-Live-1 Wins

Full-Duplex Beats Turn-Based Where It Counts

Here’s the part the pricing table hides. GPT-Live-1 is full-duplex — it listens and speaks on the same continuous audio stream, the way a phone call actually works. You can interrupt it. It can register a “wait, no” mid-sentence and adjust instead of finishing a thought nobody wants to hear anymore. Gemini 3.8 Live’s standard tier is turn-based. It waits for you to stop talking, then responds. That’s a walkie-talkie model wearing a phone-call price tag.

For scripted, single-intent tasks — “check my order status,” “book me a 2pm” — turn-based is invisible. Nobody notices the handoff because there isn’t much to interrupt. For anything closer to an actual conversation — sales calls, complex support escalations, therapy-adjacent coaching apps, any use case where a customer talks over the agent because that’s what humans do — the turn-based lag is the first thing users complain about. It doesn’t show up in a benchmark score. It shows up in call recordings and abandonment rates.

The Standard-Tier Quality Index

Reasoning tiers aside, when you compare GPT-Live-1’s standard configuration against Gemini 3.8 Live’s standard configuration on the same Speech-to-Speech Quality Index, GPT-Live-1 comes out ahead. The composite score rewards exactly the full-duplex behavior above — lower perceived latency, fewer awkward pauses, better handling of overlapping speech. Gemini’s win at the top of the market doesn’t carry down to the tier most developers will actually ship on.

Ecosystem Continuity

Teams already built on OpenAI’s Realtime API — the line GPT-Realtime-2 established back in May — inherit GPT-Live-1 without a rebuild. Function calling patterns, tool-use conventions, and integration code carry forward. That’s a real cost saving even if it never shows up on a per-minute pricing chart, and it’s the same switching-cost logic that keeps teams on GPT-5.5 even when a cheaper model scores higher on a benchmark.

The Stuff Nobody Talks About

“Cheaper” assumes your traffic matches the pricing shape. Gemini’s per-minute math looks best on listening-heavy calls. An app where the model does most of the talking — narration, walkthroughs, long explanations — shifts the blended cost closer to GPT-Live-1’s flat rate than the headline $1.38 suggests. Run your own traffic mix before assuming the discount holds.

A one-point benchmark win is not a category win. Gemini’s 82.6% against GPT-Live-1’s 81.5% is a margin, not a gap. Citing “Gemini beats GPT-Live-1 on quality” without naming the tier means quoting the one comparison that favors Google and skipping the one that doesn’t.

Turn-based isn’t a bug Google is unaware of. It’s a known trade-off, and it’s reasonable to expect Google to narrow it — the same way GPT-Realtime-2 closed the reasoning gap the original Realtime API left open. Buying today means buying today’s architecture, not the roadmap.

Neither premium tier is cheap in absolute terms. $3.50 to $4.80 an hour sounds trivial until a contact center runs reasoning-tier calls around the clock. At that point the standard-tier discount matters more, because most call volume shouldn’t need the expensive tier at all.

How to Decide

Choose Gemini 3.8 Live if:

• Your traffic is high-volume and cost-sensitive — scheduling, order status, first-tier triage
• Sessions are listening-heavy rather than talking-heavy
• You need multimodal input (camera, screen) during a live voice session
• Language coverage beyond English and a handful of major languages matters
• You can tolerate turn-based interaction on the standard tier

Choose GPT-Live-1 if:

• The product depends on natural, interruptible conversation — sales, complex support, coaching
• You’re already built on OpenAI’s Realtime API and want continuity
• Standard-tier quality matters more to you than premium-tier benchmark bragging rights
• Call abandonment or “the AI talked over me” complaints are a real metric you track

Test Both if:

• You’re building a new voice product and haven’t committed to either
• Your workload mixes scripted and open-ended calls
• The per-minute cost delta is big enough at your volume to justify a routing layer, similar to how teams route between frontier text models by task difficulty today

The Bottom Line

Google won the pricing argument decisively. An hour of standard-tier conversation costs less than half as much on Gemini 3.8 Live, and the Extended Thinking tier undercuts both GPT-Live-1 Sol and Grok Voice Think Fast 2.0 High while narrowly beating the former on quality. If your product is a high-volume, scripted voice workflow, that math is hard to argue with.

But “cheaper” and “better” are answering different questions here, and the title of this piece isn’t being cute about it. GPT-Live-1’s full-duplex architecture wins the standard-tier comparison that most developers will actually ship on, and it wins the interaction pattern that matters most in any conversation that isn’t fully scripted. Gemini’s one-point benchmark win lives at the top tier, where fewer teams will spend.

The honest recommendation: default to Gemini 3.8 Live for high-volume, low-complexity voice workloads where the price gap does real work on your margins. Default to GPT-Live-1 where the conversation itself is the product and users will notice — and complain about — a model that can’t be interrupted. Test both on your actual call recordings before committing either way. Benchmarks measure averages. Your users notice the calls that go wrong.

Frequently Asked Questions

What is Gemini 3.8 Live?

Gemini 3.8 Live is Google’s realtime voice AI model, priced at $0.005 per minute of audio input and $0.018 per minute of output on its standard tier, with a premium Extended Thinking tier at $3.50 per hour for reasoning-heavy calls. It supports 97+ languages, visual input during a live conversation, and background API calls without interrupting the audio stream. Its standard tier uses a turn-based conversational architecture.

What is GPT-Live-1?

GPT-Live-1 is OpenAI’s realtime voice AI model and the successor to GPT-Realtime-2, available through the OpenAI Realtime API. It charges a flat $0.05 per minute on its standard tier and offers a premium Sol-based reasoning tier at $4.47 per hour. Unlike Gemini 3.8 Live’s standard tier, GPT-Live-1 runs full-duplex — it can listen and speak on the same continuous stream, allowing natural interruptions.

Is Gemini 3.8 Live actually cheaper than GPT-Live-1?

Yes, substantially, on standard-tier pricing. An hour of two-way conversation costs roughly $1.38 on Gemini 3.8 Live versus $3.00 or more on GPT-Live-1 — more than double. The gap narrows at the premium tier, where Gemini’s Extended Thinking ($3.50/hour) still undercuts GPT-Live-1 Sol ($4.47/hour), but by a smaller percentage.

Which model sounds more natural in conversation?

GPT-Live-1, on the standard tier that most developers will use. Its full-duplex architecture handles interruptions and overlapping speech the way a real phone call does. Gemini 3.8 Live’s standard tier is turn-based, meaning it waits for a pause before responding — fine for scripted, single-intent tasks, noticeably less natural in open-ended conversation.

Does Gemini 3.8 Live beat GPT-Live-1 on quality benchmarks?

It depends on the tier compared. On Artificial Analysis’ Speech-to-Speech Quality Index, Gemini 3.8 Live’s Extended Thinking tier scores 82.6% against GPT-Live-1 Astra (Medium) at 81.5% — a narrow win for Google. Compare the two standard tiers instead, and GPT-Live-1 wins the composite score, largely on the strength of its full-duplex architecture.

How does Gemini 3.8 Live compare to Grok Voice on price?

Gemini 3.8 Live’s Extended Thinking tier at $3.50/hour is cheaper than Grok Voice Think Fast 2.0 High at $4.80/hour — about a 27% discount. Both sit below GPT-Live-1 Sol’s $4.47/hour, making Gemini the cheapest of the three premium reasoning tiers currently on the market.

Should I switch my voice product from GPT-Live-1 to Gemini 3.8 Live?

If your workload is high-volume and listening-heavy — scheduling, status checks, triage — the price difference likely justifies testing a migration. If your product depends on natural, interruptible conversation where users talk over the agent, GPT-Live-1’s full-duplex architecture is doing real work that a lower price tag doesn’t replace. Run both against your actual call recordings before deciding; benchmark averages won’t catch the specific failure modes your users will.

What’s the difference between full-duplex and turn-based voice AI?

Full-duplex audio lets a model listen and speak simultaneously on one continuous stream, so it can be interrupted mid-sentence and adjust in real time — the way a phone call works. Turn-based audio requires one party to finish speaking before the other responds, closer to a walkie-talkie. GPT-Live-1 runs full-duplex; Gemini 3.8 Live’s standard tier is turn-based.

---

Last updated: September 17, 2026. Sources: Google Gemini product blog · OpenAI — Introducing GPT-Live-1 · OpenAI Realtime API documentation · Artificial Analysis · xAI.

Related reading: OpenAI’s GPT-Realtime-2 Rewires Voice AI · ElevenLabs $500M ARR: Voice AI Goes Institutional · Mistral Voxtral: Open-Source Voice AI · Grok 4.6 vs GPT-5.6 Sol vs Claude Fable 5 · Google AI Pro vs Ultra Pricing · Gemini Spark Is Live