Gemini 3.1 Flash Live API
Google released Gemini 3.1 Flash Live API supporting real‑time voice agents with bi‑directional streaming, 128K context windows, and multimodal inputs—aimed at production voice systems and live agent experiences. The API is pitched as a backbone for real‑time conversational applications and embedded voice agents. (x.com)
Google published Gemini 3.1 Flash Live on March 26, 2026 and opened a developer preview for the model via the Gemini Live API in Google AI Studio. (blog.google) Developers migrating from the prior native-audio preview must change the model string from gemini-2.5-flash-native-audio-preview-12-2025 to gemini-3.1-flash-live-preview, and the model’s documented token limits are 131,072 input tokens and 65,536 output tokens. (ai.google.dev) Gemini 3.1 swaps the older thinkingBudget for a thinkingLevel parameter with settings (minimal, low, medium, high) and a default of minimal to minimize latency, and the Live API’s BidiGenerateContentServerContent events can now contain multiple content parts (for example audio chunks plus transcript) so client code must process all parts per event. (ai.google.dev) Google published benchmark figures showing Gemini 3.1 Flash Live scoring 90.8% on ComplexFuncBench Audio and 36.1% on Scale AI’s Audio MultiChallenge with “thinking” enabled, and the team says the model supports real-time conversations in more than 90 languages. (blog.google ) All audio outputs from Gemini 3.1 Flash Live are embedded with an imperceptible SynthID watermark that Google says can be detected via its SynthID detector to verify AI-generated audio. (blog.google) (deepmind.google) Google’s launch materials include product demos from partners such as Stitch and Ato, and reporting plus corporate releases indicate enterprise pilots and integrations with The Home Depot and Verizon for customer-experience workflows. (blog.google) (corporate.homedepot.com) (cnet.com)
