Google finally gave Gemini a brain and a voice—but can it pass a bank test?

Google released two new native speech-to-speech models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Unlike the old school, clunky pipelines that daisy-chained speech recognition, language processing, and synthesis, these models process audio directly. This natively eliminates the lag and "what did you say?" confusion that made previous voice bots sound like they were operating on dial-up.

The headline feature is asynchronous function calling. The AI no longer freezes in silence while it fetches data from an API; it now keeps talking, even giving status updates like a polite human. It handles visual context on the fly and switches between 97 languages mid-sentence without missing a beat. Everything coming out of the system is stamped with SynthID, because apparently, we need to watermark our own hallucinations now.

Despite the high-end benchmarks, the real-world application shows a 35.1% success rate on complex banking tasks. That means if a customer is trying to verify a wire transfer, the bot is essentially flipping a coin and hoping for the best. Access is restricted to Google AI Studio, Gemini API, and enterprise integrations like Salesforce and ServiceNow, with no option for self-hosting.

The math is simple: at roughly a dollar per hour of talk time, the human support center is officially doomed. But when the price of an error involves a customer's life savings, a 35% success rate feels like a spectacular way to bankrupt a business while appearing futuristic. Efficiency is great, but until the AI stops guessing at bank accounts, maybe keep the human in the loop.

Source: MarkTechPost