OpenAI’s voice model doesn’t think. That’s the point.
The New Stack Amanda Caswell ● Covered by 6 sources
Google and OpenAI both just tackled voice-agent lag, but in opposite ways. One keeps reasoning inside the call; the other pushes it behind the scenes.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Voice agents have a simple problem: they stall the moment they need to do real work. In the span of five days, Google and OpenAI each shipped a fix, and they point in almost opposite directions.
On Tuesday, Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking through the Gemini API and Google AI Studio, just five days after OpenAI released GPT-Live-1. Both are meant to let a voice agent keep talking while work happens in the background. The difference is where the thinking lives.
Google keeps speech, reasoning, and tool execution inside one stateful session. If a function is set to NON_BLOCKING, Gemini can keep chatting while it waits on a tool, ask a follow-up, or give an update, then continue once the result arrives. Developers can set reasoning effort to low, medium, or high per request. Standard Gemini 3.8 Live drops the extended reasoning step to save latency and tokens, and the same model also powers Gemini Live in the consumer app.
OpenAI takes the split-path approach. GPT-Live-1 handles the live conversation, while a backend reasoning model does the heavier work. That backend can be GPT-6 Astra, a lighter model like Luna, or even a third-party option. The upside is responsiveness; OpenAI puts turn-taking latency at around 800 milliseconds. The downside is that developers have to coordinate the whole dance themselves, passing context through sideband channels and deciding what happens while background jobs run.
Both systems still have to deal with the ugly reality of interruptions. If a user changes course halfway through, work may still be running in the background, and someone has to clean it up. Google says Gemini handles background noise, heavy accents, and unexpected interruptions better than competing models. It also says Extended Thinking is better for the messy conditions voice agents face outside a demo. Pricing pushes in different directions too: Gemini 3.8 Live starts at $0.005 per minute of audio input and $0.018 per minute of output, while GPT-Live-1 charges $0.05 per voice minute for the front-end voice layer alone, with reasoning and tool use billed separately.
The benchmarks don’t settle the fight, because the stacks aren’t the same. Google points to a Speech-to-Speech Quality Index score of 82.6 for Gemini 3.8 Live Extended Thinking, plus task completion rates of 68.6% on τ-Voice and 35.1% on Sierra’s τ-Voice-banking benchmark. OpenAI points to 86.2% Pass@1 on Tau3’s spoken customer-service test when GPT-Live-1 is paired with GPT-6 Astra at medium reasoning effort, and says it beat GPT-Realtime-2.1 by 30 percentage points on Full Duplex Bench. Different tests, different plumbing, different answers. That’s the real story.
My take — AI-written commentary, not fact-checked reporting
The cleanest voice model is the one that admits it doesn’t want to think in public. Google’s split may be messier for developers, but it’s honest about where the complexity lives. The broader pattern is clear: every vendor wants to sell intelligence, and every developer still gets stuck wiring the plumbing.
Read more about this at: The New Stack