Gemini 3.8 Live costs 1.38 dollars an hour. GPT-Live-1 lands closer to 5.

1.38 dollars an hour against roughly five. That is the gap between two speech to speech models released in the same week, doing broadly the same job.

OpenAI shipped GPT-Live-1. Google followed five days later with Gemini 3.8 Live on 15 September. Both target voice agents. The demos look similar. The bills do not.

The Gemini 3.8 Live pricing side

Google charges 0.005 dollars per minute of audio in and 0.018 dollars per minute out. For a conversation with a normal talking ratio, that works out at about 1.38 dollars per hour.

The architecture explains part of the price. Gemini 3.8 Live holds a continuous full duplex conversation across 97 languages and executes tool calls while it is still speaking, rather than pausing to think between turns. The standard tier skips its extended reasoning step on purpose, specifically to hold latency and token cost down.

Where the OpenAI number comes from

GPT-Live-1 quotes 0.05 dollars per minute, which sounds close enough to Google once you round. It is not the price you pay.

OpenAI splits the job in two. GPT-Live-1 handles the real time conversation. A separate backend reasoning model does the actual thinking, and it is billed separately. Put both on the same invoice and an hour of conversation costs somewhere between 4.47 and 5.83 dollars depending on how much reasoning the workload triggers.

The split has a second cost that does not appear on the bill. Orchestration between the conversational model and the reasoning model lives in your application layer, which means you own the state machine, the interruption handling and the failure modes. That is engineering time on top of inference spend, and it is the same architectural trade covered in the GPT-6 Astra benchmarks and pricing breakdown.

Latency is measured two different ways

The 0.798 second figure attached to GPT-Live-1 comes from Full Duplex Bench. That is a benchmark measurement, not API time to first audio, and the two are not interchangeable. Measured the way a developer would experience it, the number runs 1.24 to 1.34 seconds.

Neither model is slow enough to break a conversation. The point is narrower: when two vendors publish latency figures from different measurement methods, comparing them directly produces a result that is wrong by roughly 60 percent.

Why the gap decides deployments

Voice agents have been almost ready for about three years, and the blocker was never quality. It was cost per minute at the volumes a contact centre actually runs. A team handling 10,000 hours of calls a month is looking at roughly 13,800 dollars on Gemini against 44,700 to 58,300 dollars on the OpenAI stack for comparable work.

That is the kind of difference that settles a procurement decision before anyone opens a quality evaluation, which is a pattern worth reading alongside Google's recent Gemini release cadence and the Siri and Gemini arrangement now in public beta.

So: the headline per minute prices are within rounding distance of each other, and the all in costs are four times apart. Check where the reasoning happens before you check the rate card.