Google Launches Gemini 3.8 Live, Undercutting OpenAI's GPT-Live-1 on Price by Up to 70%

Google DeepMind released two new audio models on September 15, 2026: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, both available through the Gemini API and Google AI Studio. The launch directly targets OpenAI's GPT-Live-1, undercutting it on price by a wide margin.

Pricing gap

Google charges $0.005 per minute for audio input and $0.018 per minute for audio output. OpenAI's GPT-Live-1 costs $0.05 per minute. Google puts the cost of a one-hour voice conversation at roughly $1.38, compared to at least $3.00 on OpenAI's model — a difference of more than 50 percent.

Benchmark position

The Extended Thinking variant ranks first on the Artificial Analysis Speech-to-Speech Leaderboard with a score of 82.6 percent, ahead of OpenAI's latest GPT-Live-1 models, according to Google. Artificial Analysis is an independent benchmark provider, though Google is the one citing the placement in its announcement.

Capabilities

Gemini 3.8 Live is built for voice agents that need to do more than converse. According to Google, the model can:

• Make API calls in the background while a conversation continues
• Process visual input alongside audio
• Keep talking while performing other tasks
• Operate in more than 97 languages

Google has published sample applications on GitHub for developers building on the new models.

What's missing

GPT-Live-1 uses full duplex audio, letting the model listen and speak simultaneously for more natural back-and-forth conversation. Gemini 3.8 Live's architecture does not appear to match this. Based on available demos, OpenAI's model also produces more natural-sounding speech. Google's model prioritizes cost and throughput over conversational fidelity — a pattern consistent with Google's prior audio and video model releases, which have generally traded some quality for significantly lower prices.

Neither company has published independent, third-party verification of naturalness or latency claims beyond the Artificial Analysis leaderboard score, so the benchmark result should be read alongside — not instead of — hands-on testing.

What this means

This is a pricing play more than a technical leapfrog. Google is betting that most voice-agent use cases — customer service bots, background API-calling assistants, multilingual support tools — don't require GPT-Live-1's full-duplex naturalness, and that a 3-4x cost reduction will win developer adoption regardless. For high-volume deployments, the per-minute pricing gap compounds fast: a company running 10,000 hours of voice traffic a month would pay roughly $13,800 with Gemini 3.8 Live versus $30,000+ with GPT-Live-1. The tradeoff is real, though — teams building consumer-facing voice products where conversational feel matters most may still find OpenAI's model worth the premium. Expect OpenAI to respond either with a price cut or a cheaper tier of GPT-Live-1 in the coming months.