Gemini 3.8 Live & Extended Thinking: Complete Guide, Pricing, and Availability

Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026 — its most advanced real-time audio models, generally available on the Gemini API the same day. Gemini 3.8 Live targets scale, speed, and cost efficiency at $0.005 per minute of audio input and $0.018 per minute of output. Extended Thinking is a separate model that reasons and speaks in parallel, narrating its progress through multi-step tasks without dropping the conversational thread. Both are documented as Stable with 131,072-token input and 65,536-token output windows.

This guide follows the official Google announcement, the developer-focused blog post, the Gemini API model pages, the official changelog, and the DeepMind model card. It was checked on September 15, 2026, the day of the announcement.

What Google launched

The launch tweet from @GoogleAI frames the pair as “our most advanced Gemini Audio models yet”: models you can speak with, collaborate with, and hand tasks to. The company’s main @Google account added the production framing — “natural, production-ready voice applications.”

The second tweet in the thread carries the availability matrix, and it matters because the two models do not ship to the same surfaces. Three independent Google properties — the tweet thread, the blog posts, and the API changelog entry dated September 15 — agree on the launch date and the model IDs.

Two models, two jobs

Gemini 3.8 Live is the low-latency workhorse. Google’s positioning: built for “scale, speed, and cost efficiency.” It handles mid-sentence interruptions, switches languages mid-conversation (Google’s launch material says 97 languages; more on that number below), and understands visual context — point your camera at a broken bike chain or a leaky pipe in Search Live and it walks you through the fix step by step.

Gemini 3.8 Live Extended Thinking is a separate model with its own endpoint, not a toggle on 3.8 Live. It reasons and speaks in parallel: while working through a multi-step request like planning an event, it narrates its progress with short conversational fillers so the line never goes dead. The API contract makes this explicit — clients track an interaction_status signal (IN_PROGRESS or IDLE) instead of assuming silence means the model is done. Reasoning depth is configured with thinking_config.thinking_level set to low, medium, or high.

The cleanest way to remember the split: 3.8 Live answers you instantly; Extended Thinking works on your problem out loud. Notably, plain 3.8 Live rejects thinking_level entirely — sending it errors out. If you want thinking, you switch model IDs.

Benchmark snapshot

Google’s announcement leads with Extended Thinking’s claimed lead on the third-party Artificial Analysis Speech-to-Speech Quality Index, plus voice-agent evaluations. All numbers below are Google-reported; the index methodology belongs to Artificial Analysis, and Google chose the comparison configurations.

Two rows need honest framing. First, the Speech Agent Arena row is where plain 3.8 Live appears: Google reports it second on that leaderboard (press coverage adds a 76.0% index score for it, and a computed cost of about $0.84 per hour of input audio vs $3.50 for Extended Thinking). Second, Extended Thinking’s 82.6 is a large jump from the June 2026 state of the same index, when GPT-Realtime-2 led at 77.2% and the previous Gemini 3.1 Flash Live Preview sat at 69.5%. Google self-reports all of these; independent verification will come from the index maintainers, not from Google.

Pricing and context

Both 3.8 models share a single pricing row on the official Gemini API pricing page — there is no premium for Extended Thinking in Google’s rate card. The per-minute figures below come from the pricing page and match the developer blog’s footnote exactly.

Context for scale: a two-minute voice exchange with average turn-taking costs roughly a cent of input and under four cents of output on the paid tier. Free-tier usage is flagged “used to improve our products: Yes,” so production traffic belongs on the paid tier. Note that thinking tokens are billed as output — Extended Thinking’s parallel reasoning consumes output tokens even when the model is not speaking.

Where each model runs

The rollout splits cleanly, and mixing the lists up is the most common mistake in early coverage. The DeepMind model card lists the distribution for each model separately:

One maturity note worth copying exactly: the launch tweet says “public preview in the Gemini API,” while the API changelog entry dated September 15 says “generally available (GA)” and both model pages list the IDs as Stable. The Live API interface itself still carries a Preview banner in the docs. The accurate synthesis: Stable/GA model IDs on a Live API surface that remains in preview. Vertex AI availability is claimed on the DeepMind model card, but the Vertex documentation pages for these models returned 404s at publication time — treat Cloud availability as announced-but-not-yet-documented.

The Extended Thinking protocol

The genuinely new engineering surface is how Extended Thinking communicates state. On previous Live models, a turnComplete event meant the model was idle and safe to send the next prompt. On Extended Thinking that assumption breaks: the model can complete a turn’s speech, then keep working — calling tools, reasoning — and speak again. Clients must instead track interaction_status: IN_PROGRESS while reasoning or tool calls are pending, IDLE when the model is truly ready for input.

Two more contract changes matter for builders. Tools on Extended Thinking must be declared NON_BLOCKING — a blocking declaration returns a hard error, because a tool call that freezes the session defeats the point of narrated parallel work. And the model uses “conversational fillers” — short spoken placeholders like “let me check that” — to keep the dialogue alive while it works, which is the user-facing product of the interaction_status machinery.

Migrating from 3.1 Flash Live

The official migration section lists what breaks when you swap gemini-3.1-flash-live-preview for gemini-3.8-live. The full list:

• Model string: models/gemini-3.8-live in the WebSocket setup message.
• No thinking_level on 3.8 Live: omit thinking_level/thinking_config from session setup entirely — it is unsupported on this model (Extended Thinking takes low/medium/high; MINIMAL is not accepted there).
• Async function calling is the default: behavior: NON_BLOCKING. Blocking mode remains for compatibility via behavior: BLOCKING; function scheduling (SILENT, WHEN_IDLE, INTERRUPTED) is supported on 3.8 Live but not on Extended Thinking.
• send_client_content works all session long with explicit roles, replacing 3.1’s initial-history seeding; turn_complete: true now unconditionally interrupts active generation.
• Proactive audio is permanently on: setting it to false returns an error.
• Affective dialogue is removed: strip enable_affective_dialog.
• Turn coverage default changed to include all video frames — control frames explicitly if you want to manage context and cost.
• Audio is the only output modality: request output-audio transcription if you need a text transcript of what the model said.

Getting started

The Live API is a stateful WebSocket session: audio in as raw 16 kHz PCM, audio out as 24 kHz PCM, video as JPEG frames capped at 1 FPS. The minimal setup message is just the model string; voice, tools, VAD sensitivity, and transcription knobs hang off the session config. Google’s recommended production path for client-side apps is ephemeral tokens rather than shipping a raw API key.

// WebSocket setup message (minimal)
{
  "setup": {
    "model": "models/gemini-3.8-live",
    "response_modalities": ["AUDIO"]
  }
}

// Extended Thinking: separate model ID + thinking level
{
  "setup": {
    "model": "models/gemini-3.8-live-extended-thinking",
    "response_modalities": ["AUDIO"],
    "thinking_config": { "thinking_level": "medium" }
  }
}

The fastest paths: try it in AI Studio (ai.studio/live — select “Stream”), clone Google’s Live API example apps, or watch the official walkthrough:

Both models derive from Gemini 3 Pro with a documented knowledge cutoff of January 2025, and all generated audio carries SynthID watermarking. Built-in partners named in the announcement include Agora, Fishjam, LiveKit, Pipecat, Vercel, and Vision Agents — Vercel’s AI Gateway, for example, already exposes the models as google/gemini-3.8-live and google/gemini-3.8-live-extended-thinking.

Caveats

All benchmark numbers are vendor-reported. Google chose the competitor configurations for the Speech-to-Speech index comparison. The index itself is third-party (Artificial Analysis), but the 82.6 figure has not been independently republished by the index maintainers at the time of writing.

The language count is not settled. The tweet and both blog posts say 97 languages; the Live API documentation still says 70 supported languages and the models index page (last updated September 4) does not list the 3.8 models at all. Treat “97” as launch-material marketing until the API reference catches up.

No numeric latency is published. Google’s materials say “low-latency” and “near real-time” in prose only. There is no official milliseconds or time-to-first-audio figure for either 3.8 model — the 2.98 s figure circulating in coverage belongs to the older 3.1 Flash Live Preview. Live-specific rate limits are likewise unpublished; the rate-limits page (updated September 2) predates the launch.

Use it if

Use Gemini 3.8 Live if you are building a real-time voice product — support lines, voice agents, interactive assistants — and the $0.005/$0.018 per-minute rate with barge-in, visual context, and async function calling covers your loop. It is the natural default for high-volume conversational traffic.

Use Extended Thinking if your voice agent has to do things mid-conversation: look up records, call APIs, plan multi-step work — and the user should hear that it is working rather than hear silence. Budget for the extra output tokens its reasoning consumes, and design your client around interaction_status, not turnComplete.

Hold off if your deployment targets Vertex AI and you need documented Cloud support today — the Vertex docs for these models do not resolve yet. And if you depend on affective dialogue or blocking tools in a Live session, this generation removes them; plan the migration before switching model strings.

FAQ

Sources

Sources checked September 15, 2026:

• Introducing Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking — Google blog
• Build real-time voice applications with Gemini audio models — Google blog (developers)
• Gemini 3.8 Live model page — ai.google.dev
• Gemini 3.8 Live Extended Thinking model page — ai.google.dev
• Gemini API pricing — ai.google.dev
• Gemini API changelog, September 15, 2026 entry — ai.google.dev
• Thinking in the Live API guide — ai.google.dev
• Gemini 3.8 Audio model card — Google DeepMind
• @GoogleAI launch thread (tweet 1) and availability tweet (tweet 2)
• @Google corporate launch post

Related reading: Gemini 3.8 Flash complete guide and Gemini 3.8 Live API streaming guide.