Gemini 3.8 Live Pushes Voice AI Toward Stateful Agent Runtimes
Background reasoning
Gemini 3.8 Live Extended Thinking can continue spoken interaction while reasoning and running asynchronous tools in the background.
New lifecycle
Developers should track `interaction_status` and wait for `IDLE` instead of treating `turnComplete` as the end of work.
Audio costs
Paid-tier Live API audio is listed at $0.005 per input minute and $0.018 per output minute, with output pricing including thinking tokens.
Google’s new Gemini 3.8 Live family makes voice AI engineering less like building a streaming chatbot and more like operating a real-time, stateful agent runtime. The standard Gemini 3.8 Live model is positioned as the default for low-latency voice experiences. Gemini 3.8 Live Extended Thinking adds configurable background reasoning for sessions that require multi-step planning or slower tool use during spoken interaction.1
The most important change is lifecycle semantics. In Extended Thinking sessions, a spoken turn can appear complete while the model is still reasoning or waiting for asynchronous tools. Google’s documentation says developers should not treat turnComplete: true as proof that the session is idle. Instead, clients must monitor interaction_status and return the interface to an idle or listening state only when the status is IDLE.13
That design reflects a broader shift in voice agents. Traditional voice UX has been dominated by turn-taking: the user speaks, the model thinks, a tool runs and the model responds. Extended Thinking changes that sequence by allowing the model to continue the interaction with conversational fillers while it plans and calls tools in the background.1 Product teams can use that capability to reduce awkward silence, but they also inherit new obligations around orchestration, interruption handling, observability and user trust.
Google’s model-selection guidance draws a practical line between the two variants. Gemini 3.8 Live is recommended for low-latency voice agents, direct commands and fast tool execution. Gemini 3.8 Live Extended Thinking is recommended when the agent must evaluate complex data, plan across multiple steps or call tools that may take several seconds.1
That distinction matters because voice products are judged on two latency curves, not one. The first is time to first audio: how quickly the system acknowledges the user and keeps the conversation alive. The second is time to final answer: how long it takes to complete reasoning, tool calls and any grounded response. Engineering analyses of the launch emphasize that teams should test those separately because a model that starts speaking quickly may still be doing substantial work before it can provide a final, committed answer.4
For simple voice flows, such as controlling a device, answering a known FAQ or routing a support request, the standard Live model may be easier to operate. It supports both synchronous and asynchronous tool patterns, and its simpler lifecycle makes it easier for developers to map a model response to a user turn.1 For more complex flows, such as travel planning, diagnostics or tutoring, Extended Thinking offers a more natural interaction pattern by keeping speech active while reasoning proceeds.1
turnComplete is no longer enoughThe central engineering implication is that voice clients need a richer session state machine. In a conventional turn-based implementation, a frontend can often wait for a completion marker, switch the microphone back to listening and clear any “thinking” indicator. With Extended Thinking, that approach can create race conditions. The client may reopen input or mark a task as finished while the model is still running tools or preparing a final answer.13
Google’s guidance says developers adopting Extended Thinking should track interaction_status on incoming server messages. IN_PROGRESS indicates that reasoning or tool execution is still underway. IDLE indicates that the session can safely be treated as ready for the next input.1
That small API detail has large product consequences. Interfaces may need separate states for listening, speaking, thinking, waiting on a tool, recovering from tool failure and ready-for-interruption. Backend systems need correlation IDs for tool calls, response fragments and final answers. Analytics should distinguish among first-audio latency, tool latency, final-answer latency and abandoned sessions. Without that instrumentation, teams may misread a voice agent as responsive because it speaks quickly, even when task completion is slow or unreliable.
Extended Thinking also changes the tool-calling contract. Google’s documentation says function declarations for the Thinking model must use behavior: "NON_BLOCKING"; synchronous blocking tools return an error.15 In practice, developers need to treat external systems as asynchronous participants in the conversation, not hidden subroutines that pause the model until data returns.
That has several implementation effects. Tool calls need timeout policies, cancellation behavior, retries and partial-failure handling. A booking agent that searches flights, hotels and loyalty balances cannot assume every API returns in the same time window. A support agent that checks logs and account state needs a plan for stale, conflicting or missing tool responses. If the voice model continues speaking during those calls, the application must also avoid overpromising before the tool result is available.
For users, the best experience may involve explicit uncertainty: “I’m checking that now” is safer than a confident answer that later needs correction. For developers, the key is to design the spoken layer and tool layer together. The model can keep the interaction warm, but the application still needs deterministic policies for what happens when a tool fails, times out or returns after the user has interrupted.
The Live API pricing table makes production planning more complicated than a single token-rate calculation. For Gemini 3.8 Live and Extended Thinking, Google lists paid-tier input prices of $0.75 per million text tokens, $3.00 per million audio tokens or $0.005 per minute of audio, and $1.00 per million image/video tokens or $0.002 per minute. Output is listed at $4.50 per million text tokens and $12.00 per million audio tokens or $0.018 per minute, with output prices including thinking tokens.2
That pricing model encourages teams to measure the full session, not only the final answer. Long conversational fillers, retries, repeated tool calls and unnecessary background reasoning can all affect cost. Grounding also has separate economics: Google lists 5,000 free Google Search grounding requests per month across Gemini 3.x models, followed by $14 per 1,000 requests.2
Session limits also shape architecture. Google’s Live API documentation says audio-only sessions are limited to 15 minutes, audio-plus-video sessions to 2 minutes, and native audio output models have a 128,000-token context window.1 For long-running agents, developers need session resumption, summarization and context lifecycle management rather than assuming a single uninterrupted WebSocket session can carry the full experience.
The move from stop-and-wait assistants to continuous spoken interaction creates new UX questions. When the model is still thinking, should the user be able to interrupt? If a tool returns after an interruption, should the agent resume the previous task, discard it or ask for confirmation? If the system speaks fillers while waiting, how often can it do so before users perceive it as stalling?
Implementation-focused commentary around the launch frames Extended Thinking as most useful when speech must continue during multi-step reasoning, especially in WebSocket-based real-time systems that also need secure session setup, short-lived tokens and event handling.6 That reinforces the runtime-systems angle: the model is only one component. Production voice agents also require transport reliability, authentication, backpressure management, observability and graceful degradation.
The product tradeoff is not simply “faster” versus “smarter.” It is whether the user task benefits from an agent that can keep talking while background work continues. In some cases, the right answer is a faster, simpler model that completes direct commands with minimal overhead. In others, it is a more stateful agent that can coordinate tools, reason over intermediate results and manage a longer conversation.
Teams evaluating Gemini 3.8 Live Extended Thinking should start by redesigning their state model. A robust implementation should treat interaction_status as the source of truth for whether the session is still active, require all tools to be non-blocking and record separate metrics for initial speech, tool completion and final response.14
They should also test adverse cases, not just happy paths. That includes slow tools, duplicate tool responses, dropped WebSocket connections, user interruptions during background reasoning, late-arriving results and conflicting data from multiple APIs. The more natural the voice interface becomes, the more important it is for the system to be explicit about what it knows, what it is still checking and what failed.
Gemini 3.8 Live Extended Thinking points toward a new baseline for voice AI: agents that do not simply wait their turn, but maintain a live session while reasoning and acting. That can make spoken software feel more fluid. It also means developers must build voice products as distributed, asynchronous systems where latency, state and failure handling are first-class design constraints.
Anthropic’s new financial-advisor product puts Claude inside wealth-management workflows while leaving sensitive client records in partner systems. The launch is an early test of whether MCP-style connectors, enterprise controls and human approval can make vertical AI agents viable in regulated industries.
NIST IR 8587 shifts token theft prevention from post-breach cleanup to design-time controls for SSO, federation, API access and workload identity. For cloud security teams, the practical mandate is to protect signing keys, shorten credential lifetimes, validate every token path and monitor lifecycle events across providers and customers.
Cisco used .conf26 to expand Splunk’s Agentic SOC Workforce across detection engineering, threat hunting, investigation, response and policy governance. The move advances AI-assisted work allocation in the SOC, but Cisco is positioning the system around telemetry, guardrails and analyst approval rather than fully autonomous defense.
Stanford-led researchers say Paper2Agent can convert scientific papers, code and data into AI agents that answer questions, reproduce analyses and collaborate with other paper agents. The larger claim is infrastructural: papers could become executable, attributable research objects, but the system still depends on usable code, author-supplied context and human oversight.
Extended Thinking
Google’s term for a Gemini Live mode that can perform configurable background reasoning during a real-time voice session.
`interaction_status`
A Live API session field that indicates whether the model is still in progress or idle, used to manage client state.
Non-blocking tool call
An asynchronous function call that does not pause the spoken interaction while the external tool runs.
Time to first audio
A voice-agent latency metric measuring how quickly the system begins speaking after user input, separate from time to final answer.
Comments