Five days. OpenAI put GPT-Live-1 into the API on September 10 with full-duplex listening, and Google’s answer arrived on September 15: two voice models, one of them now sitting at the top of Artificial Analysis’ Speech-to-Speech Quality Index. The voice-agent race has a response cycle measured in days, which tells you how seriously both companies want this market.
It also means builders now have a real choice for the first time. Here’s what shipped, and how I’d pick between the two if I were putting a voice agent into production this quarter.
What Google shipped
Two models, aimed at different jobs. Gemini 3.8 Live is the scale-and-cost option. Gemini 3.8 Live Extended Thinking is the one for hard problems, and it took the #1 spot on the Speech-to-Speech Quality Index with a score of 82.6.
Per Google’s announcement, Extended Thinking leads agentic task completion at 68.6% on τ-Voice and 35.1% on Sierra’s τ-Voice-banking variant, and scores 97.7% on Big Bench Audio. The plain 3.8 Live holds second place in the Speech Agent Arena while staying cheap enough to run at high volume. On ServiceNow’s EVA-Bench, both models push the Pareto frontier for complex voice workflows, which is a fancy way of saying you get better completion rates without paying proportionally more for them.
One caveat worth stating plainly: most of these numbers are vendor-reported. Google picked the benchmarks and ran them. Treat the rankings as directional until third-party evals pile up.
The line between a voice interface and a voice agent
Plenty of models can talk. The dividing line in 2026 is what happens when the conversation gets complicated.
Extended Thinking reasons while speaking. It handles multi-step problems mid-conversation, runs background tools without going silent, uses real-time visual context, absorbs interruptions, and switches languages on the fly. The background-tool piece matters most. A voice interface answers; a voice agent can say “give me a second, checking that,” go run a tool, and come back without you having to call again.
That capability ships in the Gemini API and AI Studio today, plus Search Live, the Gemini app, and Workspace surfaces like Docs, Gmail, and Keep. Enterprise deployment is in private preview under Gemini Enterprise. GPT-Live-1’s counter-advantages: it reached the API first, and its full-duplex design lets it listen while it talks.
How to pick: four questions
1. What completion rate do you actually need?
If your agent books meetings or reads out order status, the plain 3.8 Live tier is probably enough, and its price point is built for that volume. If the agent has to resolve multi-step problems (banking workflows are the canonical example, and τ-Voice-banking’s 35.1% top score shows how hard those still are), you want Extended Thinking. The gap between 68.6% general completion and 35.1% banking completion is your reality check: voice agents are strong on general tasks and still mediocre on domain-locked ones.
2. What does an hour of input audio cost you?
Benchmark wins are nice; unit economics decide deployments. Price both options per hour of input audio at your real concurrency, instead of per API call. Long-running agents re-read context constantly, so ask each vendor what a cache-heavy workload looks like at 3am on a Tuesday before you look at headline rates.
3. Where does it need to live?
Google’s rollout touches surfaces OpenAI doesn’t have: Search Live, Workspace apps, distribution through the Gemini app. If your users live inside Google’s ecosystem, that surface coverage is worth real money. If you’re building a standalone product on the API, both vendors are one integration each, and GPT-Live-1 got there first.
4. Which benchmarks do you trust?
The eval names that matter for voice have changed, and procurement teams already know it. When someone shows you a voice benchmark now, the question is whether it’s τ-Voice (agentic completion), EVA-Bench (enterprise workflow cost-performance), or the Speech Agent Arena (open comparison). Anything else might be a home-field test the vendor built to win.
What to do now
Three moves this week, roughly two hours total:
- Pull both APIs into a spike project and run ten real calls from your actual workflow, five simple, five that require a tool call mid-conversation. The tool-call-under-interruption case is where they’ll diverge.
- Compute cost per resolved task rather than per call. A cheaper model that fails 30% of tasks and forces retries is the expensive one.
- Write down your latency and interruption requirements before you look at another benchmark table. It kills most of the debate early.
And keep one eye on the cadence. Five days between OpenAI’s move and Google’s response means whatever you pick this month is a rolling decision, not an architecture. Build the swap behind an interface you control.
The voice-agent market spent years as a demo category. Two credible API options and a competition cycle measured in days: that’s a production market now. The winners will be the teams that measure cost per resolved task instead of arguing about which model sounds more natural.
