Gemini 3.8 Live vs GPT-Live-1: Google Undercuts OpenAI's Voice Model by 2x — Is It Worth Switching?

2026-09-16 · 4 min read

geminimodel-releasevoicecomparison

What's new

On September 15, 2026 Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two speech-to-speech models available through the Gemini API and Google AI Studio. They take audio in and produce audio out, can call tools and APIs mid-conversation, accept visual input, and support 97+ languages. Google has published sample applications on GitHub.

The headline is price. Google charges $0.005 per minute of audio input and $0.018 per minute of output, which works out to roughly $1.38 for an hour of conversation. OpenAI's GPT-Live-1, released five days earlier, is priced at $0.05 per minute, or about $3.00 per hour. That's a little over a 2x gap.

On quality, the Extended Thinking variant ranks first on the Artificial Analysis Speech-to-Speech Leaderboard at 82.6%, ahead of GPT-Live-1. But The Decoder's reviewer notes the trade-off plainly: OpenAI's model is full duplex (it listens while it speaks, so it can be interrupted naturally), and it "sounds better" in practice. Google appears to have optimised for price over voice quality.

Why a voice model costs by the minute, not the token

Voice models are billed differently from the text models most people use day to day. A text model charges per token; you pay for what you type and what it writes. A speech-to-speech model charges per minute of audio, whether or not anyone is saying anything useful. That's why the "$1.38 vs $3.00 per hour" framing matters: for a call-centre bot or a language-tutoring app, the meter is running the entire session, and a 2x price difference compounds across thousands of hours.

It also explains why there are two variants. Plain 3.8 Live is the cheap, fast one for simple back-and-forth. Extended Thinking is the one that reasons before it answers — better on the leaderboard, but with more latency between your sentence and its reply, which in a voice product is felt immediately.

Is it on AIWITH.CHAT?

No, and to be straightforward about it: neither Gemini 3.8 Live nor GPT-Live-1 is a chat model. They are API building blocks for developers shipping voice products. AIWITH.CHAT is a text-and-image chat platform, so there is nothing to "switch on" here.

What you can use on AIWITH.CHAT is the text side of the same Gemini family — Gemini 3.1 Pro (the default) and Gemini 3.6 Flash — alongside GPT-6 Astra, Claude Fable 5.1, DeepSeek V4.1 and Kimi K3 on a single $9.9/month plan, no API key required. If your interest in 3.8 Live is really "is Google's current generation good enough to move my daily work to?", that's the question you can answer today, in a browser, for the price of a coffee.

See plans →

Is it worth switching?

Three kinds of reader, three straight answers:

You're building a voice product and already pay OpenAI per minute — yes, run a pilot now. Halving the per-hour cost is real money at scale, tool calling is supported, and the leaderboard score says the reasoning is at least as good. The thing to test is interruption handling: without full duplex, a support bot that talks over the caller will cost you more in churn than you save on the meter.
You're building a voice product where naturalness is the product (companionship, coaching, anything premium) — probably not yet. Reviewers consistently say GPT-Live-1 sounds better and handles being cut off more gracefully. Pay the 2x until Google closes that gap.
You're an individual using AI for writing, coding, research or study — this release doesn't touch you. There's no chat interface for it, and it does nothing a text model can't do faster in a browser. Keep using Gemini 3.1 Pro (or whichever model fits the task) and ignore the noise.

In one line: Gemini 3.8 Live is a price move aimed at developers, not a capability jump for end users. If you pay per minute, test it this week; if you pay for a subscription, nothing changes.