The news

Google's Gemini Audio team posted Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking on 15 September 2026 (page marked updated 17 September 2026): blog post. Two Live API model strings are documented the same week:

  • gemini-3.8-live: default low-latency voice / real-time dialogue without reasoning-induced delays; interleaved reasoning; async function calling; Live API supported; Search grounding supported.
  • gemini-3.8-live-extended-thinking: high-reasoning audio-to-audio for complex multi-step work during live voice; background reasoning plus async tool calls while streaming continuous audio.

Focus: Live Extended Thinking, and how Google is bundling the pair into Search Live, the Gemini app, and Google Workspace (Docs Live, Gmail Live, Keep Live). Keep OpenAI's GPT-Live-1 priced full-duplex API story in a separate file.

Who is bound

  • Google ships the models and the consumer / Workspace surfaces.
  • API developers on Gemini API and Google AI Studio (both SKUs rolling out "starting today" on the blog).
  • Enterprises on Gemini Enterprise (private preview; Customer Experience "coming soon"; Workspace business customers called out for Extended Thinking).
  • End users on Search Live (3.8 Live), Gemini Live, and Google AI subscription tiers for Workspace surfaces (blog: AI Pro and Ultra for Docs; all Google AI subscribers for Gmail and Keep on Extended Thinking).
  • Named platform partners on the post (Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, Vision Agents) and design partners (Salesforce, Genspark, Lumeris) are integration paths, not your SLA counterparties unless you contract them.

What's new

  1. Two-SKU Live split. Blog: 3.8 Live for scale and cost efficiency with conversational intelligence, fluid dialogue, and visual grounding; 3.8 Live Extended Thinking for high-complexity tasks with increased intelligence and multi-step reasoning. Docs put both at 131,072 input / 65,536 output tokens; September 2026 update stamp.
  2. Extended Thinking behavior. Blog: reasons and speaks simultaneously; early verbal cues ("Let me check that…"); live progress narration for multi-step background tasks; demos include sketch-to-React, multi-step bookings with asynchronous function calls, and business-plan / marketing toolkit generation by voice. Docs: only async non-blocking function calling (behavior: NON_BLOCKING); synchronous blocking returns a hard error; configure thinking_config with thinking_level low / medium / high (MINIMAL not supported); after turnComplete: true the server may still be in IN_PROGRESS for background reasoning or tools until interaction_status is IDLE.
  3. 3.8 Live migration notes (docs). From gemini-3.1-flash-live-preview: drop thinking_level / thinking_config on the non-Extended SKU; async function calling is now the default (NON_BLOCKING) with optional BLOCKING for compatibility; proactive audio permanently enabled; affective dialogue removed; turn coverage defaults include audio activity and all video.
  4. Company eval claims (blog, Google's numbers). Extended Thinking: #1 overall on Artificial Analysis Speech to Speech Quality Index at 82.6; 68.6% on τ-Voice; 35.1% on Sierra τ-Voice-banking; 97.7% on Big Bench Audio. Live: second place in Speech Agent Arena; EVA-Bench (ServiceNow) cited for Pareto balance of accuracy and conversational quality on the Live API on Gemini Enterprise Agent Platform. Those benchmarks are Google's citations, not The Frontier's re-run.
  5. Product bundling, not only an API SKU. Same post ships the models into Search Live (step-by-step troubleshooting), Gemini app (Daily Brief, inbox, to-dos), and Workspace Docs / Gmail / Keep Live surfaces. SynthID watermarking is stated for all audio from Google's AI products.
  6. Capability surface. Near real-time visual inputs; automatic detection and mid-conversation transition across 97 supported languages; tools and API calls in the background while conversation continues.

What it does not settle

  • API list price for either Live SKU is not stated on the blog or the two model docs we used. Treat pricing as UNKNOWN here (unlike the GPT-Live-1 $0.05/min front-end figure elsewhere on The Frontier).
  • Independent replication of Speech to Speech Quality Index, τ-Voice, τ-Voice-banking, Big Bench Audio, Speech Agent Arena, and EVA-Bench scores is UNKNOWN.
  • Enterprise GA dates beyond "private preview" / "coming soon" for Customer Experience and Workspace business are UNKNOWN.
  • Parity of Extended Thinking features across API vs consumer Workspace (thinking levels exposed to end users, tool allow-lists, data retention) is UNKNOWN on these pages.
  • Whether Live Extended Thinking displaces text-only Gemini thinking models for the same tasks is a product choice, not a claim settled by the launch post.

What to do now

  1. Voice-agent builders: if you need background reasoning during a live call, target gemini-3.8-live-extended-thinking and redesign clients around interaction_status (IN_PROGRESS vs IDLE) rather than treating turnComplete as idle.
  2. Migrators from 3.1 Flash Live: switch the model string to gemini-3.8-live, remove affective-dialogue and proactive_audio: false configs, and decide explicitly between default async tools and legacy BLOCKING.
  3. Workspace / Search operators: treat Docs Live, Gmail Live, Keep Live, and Search Live as the consumer proof points for Extended Thinking bundling; do not wait for a separate "voice API price" post to evaluate UX.
  4. Keep the OpenAI comparison clean: GPT-Live-1 coverage on The Frontier is a priced full-duplex API front end with backend delegation. This Google story is a dual Live SKU plus Search/Workspace distribution. Do not merge them into one "voice API week" ledger.