Gemini 3.8 Live

Gemini 3.8 Live is a genuinely novel native speech-to-speech model that collapses the traditional ASR-LLM-TTS pipeline into a single hosted API, with asynchronous tool calling that keeps the conversation flowing during background tasks -- a real architectural advance over prior-generation voice agents. Its headline cost ($0.84/hour of input audio, roughly $1.38/hour all-in by some estimates) is meaningfully lower than OpenAI's GPT-Live-1 Astra ($5.83/hour input per Artificial Analysis), which is the most credible competitive advantage. However, the benchmark lead is narrow (82.6 vs. 81.5 for Extended Thinking vs. GPT-Live-1 Astra), benchmark rankings in this space have reversed multiple times in the past year, and in blind human preference testing Google's own older 3.1 Flash Live still outranked 3.8 Live in the Speech Agent Arena. The models are hosted-only with no self-hosted option, enterprise access remains in private preview, and no independently named enterprise customers with production deployments have been disclosed beyond partnership announcements with Salesforce and ServiceNow. Enterprises should treat this as a high-priority pilot candidate given the price differential, but should hedge against rapid benchmark volatility and the absence of audited compliance documentation at launch.