語音 AI／模型 API
Gemini 3.8 Live Moves Reasoning into the Voice Background, So Clients Can No Longer Rely on turnComplete to Determine Idleness
Google has launched two native-audio models. Extended Thinking can perform background reasoning and asynchronous tool calls while continuously streaming audio. The new protocol improves concurrency for long-running voice agents, but also requires developers to rewrite session state machines and interruption handling.
On September 15, Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both are based on Gemini 3 Pro, accept text, image, video, and audio input, and produce text and native audio output. The [model card](https://deepmind.google/models/model-cards/gemini-3-8-audio/) lists a maximum input context of 128K tokens and a maximum output of 64K tokens. The models are positioned for high-volume, low-latency, real-time conversations, while Extended Thinking additionally allows the model to perform background reasoning and tool work as it streams speech.
The protocol semantics are what will truly affect existing integrations. According to the [developer documentation](https://ai.google.dev/gemini-api/docs/models/gemini-3.8-live-extended-thinking), receiving `turnComplete: true` no longer means the server is completely idle; reasoning, tool calls, or subsequent audio may still continue. Clients must track the new `interaction_status`, and work can be considered complete only when it reaches `IDLE`. Function calls support only the asynchronous `NON_BLOCKING` mode. If an application still closes the socket, releases tool context, or accepts the next mutually exclusive task after turn completion, it may lose calls or create race conditions. Developers can also set background reasoning to low, medium, or high, but cannot select minimal, and `proactive_audio` must remain enabled.
Google’s evaluations use single-attempt responses and cover Artificial Analysis, ServiceNow EVA-Bench, and Sierra τ³-Banking. Extended Thinking was tested with the high reasoning level and default sampling settings. These scores cannot be directly reduced to real-world customer-service success rates: Sierra’s [research on voice benchmarks](https://sierra.ai/uk/blog/bench-advancing-agent-benchmarking-to-knowledge-and-voice) shows that task-completion rates fall substantially across systems when noise, interruptions, and speech not directed at the agent are introduced. The next questions to watch are how much background reasoning increases time to first audio and cost, and whether applications can maintain consistent session state across reconnections, user interruptions, and non-idempotent tools.