Google's Gemini 3.8 Live Tops Speech Leaderboard at $1.38 an Hour, Undercutting OpenAI's GPT-Live-1

Google DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two audio models aimed at developers building voice applications, and both now sit at the top of the Artificial Analysis speech-to-speech leaderboard. The pricing is the sharper edge of the announcement: Google is charging $1.38 per hour of voice conversation, a rate that undercuts OpenAI's GPT-Live-1 by a wide margin. For teams weighing which model to wire into a call center, tutoring app, or voice agent, that gap changes the math on every deployment decision.

The two models differ mainly in how much reasoning they do before answering. Gemini 3.8 Live handles standard conversational turns, while the Extended Thinking variant spends more compute on harder requests, which matters for tasks like multi-step troubleshooting or spoken math. Both are available through Google's developer channels, positioning them against a crowded field of voice-first APIs rather than general text models. Artificial Analysis, which runs head-to-head evaluations across providers, places the pair ahead of the competition on speech-to-speech quality as of release.

Cost has become the central battleground in voice AI. A voice agent running eight hours a day at $1.38 per hour costs roughly $11 per day in inference, before any markup, which is the kind of number that makes always-on assistants financially plausible for smaller companies. OpenAI's GPT-Live-1 has been the reference point for natural-sounding conversation, and it remains the model to beat on pure realism. Google's answer is to win on price and benchmark placement rather than claim a subjective edge in how human the voice feels.

That realism gap is worth watching, because it comes down to architecture. GPT-Live-1 uses full duplex, meaning the model can listen and speak at the same time, which produces the interruptions, backchannels, and overlapping turns that make a conversation feel unscripted. Many speech-to-speech systems, including turn-based ones, wait for the user to stop before responding, and that pause is audible to anyone who has talked to a voice bot. Google has not claimed full duplex parity in this release, so the leaderboard win and the pricing advantage may not translate into the most natural call experience.

The release also signals how Google wants developers to build. Extended Thinking as a separate audio model suggests the company expects voice agents to handle reasoning-heavy prompts, not just chit-chat, and pricing that low is an invitation to route production traffic through Gemini rather than a smaller vendor. OpenAI now faces pressure to respond on cost without giving up the conversational quality that justified its premium. The open question is whether developers prioritize the cheaper benchmark leader or the model that sounds more human when a customer is actually on the line.

Expect the next few weeks to bring independent latency and interruption tests, since benchmark rankings rarely capture how a model behaves mid-sentence. If Gemini 3.8 Live closes the duplex gap at its current price, OpenAI's pricing on GPT-Live-1 becomes difficult to defend for high-volume use cases. If it does not, Google may win the spreadsheet while losing the demo.