Google has made Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking generally available through the Gemini API, adding a production-ready Google option for teams building real-time voice agents. The release matters less as a single model announcement than as another sign that voice AI is becoming a metered infrastructure choice, not an experimental interface bolted onto text models.
The new Gemini Live models are audio-to-audio systems exposed through Google’s Live API. Developers can use them for real-time voice applications where the model listens, reasons, and responds in speech without forcing every interaction through a text-first workflow. Google is also making the models available through Google AI Studio, which gives teams a lower-friction way to test voice behavior before wiring it into production applications.
The commercial detail is unusually important: Google lists audio pricing at $0.005 per minute for input and $0.018 per minute for output. That puts voice usage into a billing unit that operations teams can understand alongside support minutes, call-center handling time, and other service-delivery metrics. For businesses comparing voice APIs, the unit economics are now easier to model than token-only pricing, though the total bill will still depend on session length, turn-taking, tool calls, and any additional model usage around the voice session.
What changed
Gemini 3.8 Live is now a generally available model family for production voice applications, rather than a preview feature or a narrow experimental endpoint. Google’s release notes list both Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking as GA on September 15, 2026. Google describes the models as real-time audio-to-audio models for the Live API, with availability in both the Gemini API and AI Studio.
The “Extended Thinking” variant is particularly relevant for more complex voice agents. In practice, production voice systems often need to do more than answer a short question. They may need to check account state, ask clarifying questions, summarize a policy, route a support case, or decide whether to escalate to a human. A voice model that is positioned for deeper reasoning may reduce the amount of orchestration developers need to build around the model, although teams will still need to test latency and reliability under real workloads.
The release also changes the competitive landscape for multi-provider voice infrastructure. Recent voice API releases have made real-time speech a more explicit platform category. Google’s GA entry gives developers another large-provider option to evaluate for speech quality, latency, tool integration, pricing, regional availability, and enterprise procurement fit.
Why it matters for developers
For developers, the immediate work is catalog and integration hygiene. Model IDs, capabilities, supported modalities, pricing units, and routing rules need to be represented accurately in internal tooling. A team that already routes text, image, and embedding calls across providers cannot treat live voice as just another chat completion endpoint. Voice sessions have different lifecycle behavior, different failure modes, and different cost drivers.
Real-time voice applications also put pressure on observability. Developers need to know where latency is introduced, how often sessions are interrupted, whether audio input is being interpreted correctly, and how costs accrue across input and output minutes. Usage dashboards that only show token counts will miss a central part of the operational picture. For teams running an enterprise LLM API layer, Gemini 3.8 Live should be added as a distinct voice-capable model with separate analytics and policy handling.
The per-minute pricing also affects testing. Teams should not evaluate voice agents only by subjective conversational quality. They should measure average session duration, silence handling, retries, transfers, tool-call frequency, and abandonment rates. A cheaper model per minute can still be expensive if it produces longer calls, needs more clarification turns, or requires a fallback model for complex tasks. Conversely, a higher output-minute cost may be acceptable if it shortens sessions and reduces human escalation.
Business impact
The biggest affected groups are customer-support teams, sales-assistance teams, contact-center platforms, workflow automation vendors, and internal IT or HR service desks that are experimenting with speech-first interfaces. For these teams, GA status is a procurement and risk signal. It does not remove the need for testing, but it makes the model easier to consider for production roadmaps than a preview-only release.
Businesses should treat the release as a reason to update vendor comparison sheets and cost models. Voice agents combine several cost surfaces: inbound audio, outbound audio, any text or reasoning calls around the session, tool calls, logging, storage, and human escalation. Google’s published per-minute audio rates make one part of that calculation clearer, but they do not produce a full cost estimate by themselves.
Governance teams also need to revisit policy boundaries. A voice agent may handle personal data, payment-related questions, medical or legal context, or confidential enterprise information. That means teams need routing controls, redaction rules, retention policies, audit logs, and incident procedures before a voice model is broadly deployed. The model release is a capability event, but the enterprise question is whether the surrounding controls are mature enough for the use case.
For API platforms and resellers, the release creates a catalog-management task. Gemini 3.8 Live and its Extended Thinking variant need to be exposed with accurate names, supported endpoints, pricing metadata, and product constraints. Partner-facing dashboards should distinguish voice minutes from token usage so customers can attribute spend to applications, teams, or customers. Without that separation, voice adoption can blur cost ownership quickly.
What remains uncertain
The main unknowns are practical rather than factual. Google has published the availability and audio pricing, but each production team still has to test the model against its own environment. Real-time voice performance depends on network conditions, microphone quality, accents, background noise, turn-taking behavior, tool latency, and the design of the agent’s instructions.
It is also too early to know how buyers will weigh Google’s voice models against competing real-time APIs. Some teams will prioritize latency and speech naturalness. Others will care more about integration with existing Google Cloud or Gemini workflows. Larger enterprises may make the decision based on governance, procurement, data-handling terms, observability, and the ability to route traffic across providers.
The practical takeaway is straightforward: Gemini 3.8 Live gives voice-agent builders another GA option, and its per-minute pricing makes the cost model easier to operationalize. The next step for most serious teams is not a broad migration. It is a structured evaluation: add the models to the internal catalog, run production-like voice scenarios, measure minute-based spend, and decide where Google’s Live API belongs in the broader multi-model routing strategy.