Gemini 3.8 Live landed in my stack the same week I was still arguing about whether speech-to-speech belonged on every production number.
Google DeepMind announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking around 2026-09-15. Both ship through the Live API. Google named LiveKit among the day-one media partners that wrap realtime transport so teams can swap model strings without rebuilding WebRTC. Audio pricing lands at $0.005 per minute input and $0.018 per minute output, estimated from $3 and $12 per million tokens.
That price looks friendly until you treat both variants as one dial.
Two model strings, two cost shapes
gemini-3.8-live is documented as the default for low-latency voice agents and realtime dialogue without reasoning-induced delays. It supports interleaved reasoning, asynchronous function calling, and full-session client content updates. Async tool behavior is the default. You can still force blocking tools when you need them.
gemini-3.8-live-extended-thinking is the high-reasoning audio-to-audio path. It keeps streaming spoken progress while background reasoning and async tools continue. Function calling is async-only. Blocking mode hard-errors. You configure thinking_level as low, medium, or high. Minimal is not supported. Client state management changes: turnComplete: true no longer means idle. You wait for interaction_status: IDLE before assuming the turn is finished.
Google also positions these native speech-to-speech models as a streamlined alternative to cascaded STT-LLM-TTS stacks. That is an architecture claim, not a free lunch. Cascaded stacks still win when you need separate billing, separate failover, or hard voice branding. Native S2S wins when you want one session, barge-in, and fewer hop failures. The choice is cost topology, not fashion.
Where Extended Thinking belongs on a live call
I run usage-based voice for paying customers. The expensive pattern is obvious: default every session to Extended Thinking because a benchmark said it ranked first on a speech-to-speech quality index. Then watch ordinary FAQ turns pay for background reasoning they never use.
Putting Extended Thinking on the default live path is how teams burn latency and token budget while calling it intelligence.
Use 3.8 Live for greetings, authentication prompts, simple lookups, and any flow where the caller barges in constantly. Promote Extended Thinking when the caller enters multi-step booking, claims adjudication, or tool chains that actually need narrated progress while work continues. That routing decision belongs in your agent graph, not in a global env var.
LiveKit day-one support matters here as an architecture commitment. If your media plane and model plane cannot accept the same model IDs on announce week, you are stuck on last month's latency profile while your unit economics assume this week's rates. Partner listing is not the same as a changelog badge in every repo, but Google's developer post is explicit: LiveKit is in the Live API partner set for realtime deployment.
Quiet cost levers operators miss
Video frames are on by default for turn coverage. Send frames only when the product needs visual grounding. Otherwise you feed tokens into a path that never uses them.
Session state bugs on Extended Thinking are a cost bug. If your client closes after turnComplete while the server is still IN_PROGRESS, you retry tools, double-fire webhooks, or leave sockets half-open. That shows up as minutes and failed calls, not as a model quality chart.
Measure per-intent minutes, not blended ARPU. A path that looks cheap at roughly $0.023 per combined audio minute can still wreck margin if Extended Thinking is the default and half your traffic is "what are your hours." Track promotion rate into Extended Thinking the same way you track escalation to a human.
I am not dropping cascaded pipelines where Twilio or Telnyx handoff still needs discrete STT for compliance transcripts. I am saying the new Gemini Live pair forces an explicit fork: low-latency S2S for the common path, Extended Thinking as a promoted lane, LiveKit (or equivalent) as the media contract that lets you change the model string without rewriting telephony.
Ship the router first. Then celebrate the model card.