Google Launches Gemini 3.8 Live for Voice Agents: Pricing, Privacy, and Migration

*This research-based analysis covers Google's September 15, 2026 release and documentation available on September 16. DiscoverAI has not independently benchmarked the models; capabilities and pricing attributed to Google remain provider claims and can change.*

The short answer

Google released two stable Gemini Live API models on September 15: `gemini-3.8-live` for low-latency audio conversations and `gemini-3.8-live-extended-thinking` for voice interactions that need background reasoning and asynchronous tools. Both accept text, images, audio, and video, return text and audio, and expose a 131,072-token input limit and 65,536-token output limit.

Paid Gemini Developer API pricing is currently $0.75 per million text-input tokens, $3 per million audio-input tokens, $1 per million image/video-input tokens, $4.50 per million text-output tokens, and $12 per million audio-output tokens. Google's approximate audio rates are $0.005 per input minute and $0.018 per output minute. Those figures do not describe a whole call: persistent sessions repeatedly process retained context, transcriptions add tokens, and tool or Search usage can add cost.

What is the difference between Gemini 3.8 Live and Extended Thinking?

Gemini 3.8 Live is the default for immediate conversation, customer-service triage, voice search, language practice, and fast tool calls. It supports interleaved reasoning and non-blocking function calls without inserting a long reasoning pause into every turn.

Extended Thinking is intended for multi-step work that must keep the conversation alive while the model plans or waits for tools. It can speak an intermediate update, continue reasoning in the background, call a tool, and speak again. That changes the client contract: turnComplete no longer means the interaction is finished. Applications must watch interaction_status and wait for IDLE before treating the model as done.

That distinction matters for reliability. A voice interface that unlocks an account, books travel, changes a record, or sends a message needs explicit action states, idempotency, timeouts, cancellation, and confirmation—not merely fluent audio.

How much does Gemini 3.8 Live cost in practice?

Google bills Live API use by tokens. Audio input is listed at about half a cent per minute and audio output at about 1.8 cents per minute, before text, images, video, transcription, grounding, or accumulated history. A simple one-minute input plus one-minute output therefore starts near 2.3 cents under the published approximations, but a real session can cost more.

The important wrinkle is context compounding. Each turn can rebill the retained conversation, including native audio tokens. Proactive audio is permanently enabled for both 3.8 models, so input is billable while the API is listening. Input and output transcription also create text-token charges. Google recommends context-window compression to evict older material and stop long conversations from growing without bound.

Teams should measure cost per completed, accepted task across real call lengths. Track silent listening, interruptions, repeated context, tool retries, Search queries, transcriptions, transfers, and human correction. A per-minute list rate is useful for orientation, not forecasting.

What changes when migrating from Gemini 3.1 Flash Live?

For the standard model, change the model string to gemini-3.8-live, remove thinking_level or thinking_config, and review non-blocking function behavior because asynchronous execution is now the default. Proactive audio cannot be disabled, affective dialogue has been removed, and video frames are included by default when sent.

Extended Thinking requires a larger state-machine change. The client must continue receiving messages after turnComplete, distinguish in-progress from idle states, reconcile tool calls that finish out of order, and let a user interrupt safely. Run regression tests for barge-in, duplicate actions, late tool results, network reconnection, and escalation before changing a production model ID.

What are the privacy and retention boundaries?

Google says paid Gemini API prompts and responses are not used to improve its products. Free-tier content may be used for improvement. Optional developer-owned logging on billing-enabled projects can retain request and response logs for 7, 14, 28, or 55 days, and deliberately shared datasets may be reviewed and used for product development.

The Live API also has a specific retention trap: configuring session resumption can retain conversation state—including text, audio, and video—for up to 24 hours. Google says developers pursuing zero data retention should not enable that configuration. Paid use can still involve limited abuse-monitoring retention, and Google directs workloads requiring guaranteed zero data retention or enterprise agreements to Vertex AI.

For sensitive voice workflows, disclose recording and AI use, collect only necessary audio, keep credentials server-side or use short-lived ephemeral tokens, redact transcripts, restrict tools, and document deletion. A natural voice does not reduce the sensitivity of what callers say.

The verdict

Gemini 3.8 Live makes Google's real-time voice stack more production-shaped: stable model IDs, multimodal input, asynchronous tools, and a separate reasoning model give developers clearer choices than one general preview endpoint. Extended Thinking is especially notable because it lets complex tool work continue without turning the conversation into dead air.

The tradeoff is a more complicated and potentially expensive session lifecycle. Teams should choose the standard model unless a measured task benefits from background reasoning, cap retained context, and require human confirmation for consequential actions. Low latency can make a voice agent feel trustworthy; it does not make its reasoning, tool calls, or permissions trustworthy by default.