Live is two SKUs

If your RFP still says “Gemini Live” as one line item, you are already shopping wrong.

On September 15, 2026, Tom Ouyang and Malini Jaganathan (for the Gemini Audio Team) introduced two models on the same day: Gemini 3.8 Live, built for scale and cost efficiency with fluid dialogue and visual grounding, and Gemini 3.8 Live Extended Thinking, built for high-complexity tasks with increased intelligence and multi-step reasoning. The developer companion post (Alisa Fortin and Thor Schaeff, same day) puts both on the Live API in Gemini API and Google AI Studio, with a shared audio price point and a clear split in job: keep talking while tools run, or keep talking while deeper background reasoning runs too.

This review is primary-source plus same-day independent writeups. I do not have hands-on access to enterprise private preview or to every Workspace Live surface. Where Google states a benchmark or a rollout gate, I cite it as vendor-stated. 9to5Google (Abner Li, September 15) and SiliconANGLE (Mike Wheatley, September 15) corroborate the dual-model launch, the Spoken benchmark numbers Google published, and the surface split. Where docs disagree with marketing softness, I follow the docs.

What it is

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are native speech-to-speech models in the Gemini 3 series, based on Gemini 3 Pro per the DeepMind model card published September 15, 2026. Inputs: audio, images, video, and text, with up to about 128K to 131K context depending on whether you read the model card or the API tables (card says up to 128K; API docs list 131,072 input and 65,536 output tokens). Outputs: audio and text.

They are sold as the building blocks for production voice agents and as the models behind more fluid spoken experiences in the Gemini app, Google Workspace, and Search. Partner platforms named on the product blog for Live API integration include Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents. Enterprise name-drops include Salesforce, Genspark, and Lumeris (excitement quotes, not independent evals).

The dual-SKU point is not branding fluff. API model codes are separate: gemini-3.8-live and gemini-3.8-live-extended-thinking. Capabilities diverge on thinking configuration, function-calling rules, and where each model shows up for end users.

What changed

Relative to prior Live generations (the Live docs call out migration from gemini-3.1-flash-live-preview), 3.8 Live is positioned as ultra-low latency audio-to-audio with interleaved reasoning, asynchronous function calling, full-session client content updates, and built-in audio streaming. Thinking level is not supported on gemini-3.8-live. You omit thinking_level / thinking_config. Async non-blocking function calling is the default. Synchronous blocking remains available for compatibility. Proactive audio is permanently on. Affective dialogue configuration is removed.

Extended Thinking is the other product. Docs recommend it when higher background reasoning is required for complex, multi-step problem solving during live voice. It processes background reasoning and asynchronous tool calls while streaming continuous audio. Only asynchronous non-blocking function calling is supported. Synchronous blocking returns a hard error. Thinking is configured with thinking_config at low, medium, or high. Minimal is not supported. Critically for engineers: turnComplete: true no longer means the server is idle. Clients must keep listening and watch interaction_status (IN_PROGRESS vs IDLE) because reasoning or tool calls may continue after a turn looks complete.

Product surfaces also split. Per the September 15 product blog:

3.8 Live rolls out for developers in Gemini API and AI Studio; for enterprises in private preview in Gemini Enterprise (coming soon to Gemini Enterprise for Customer Experience); for everyone in Search Live.

3.8 Live Extended Thinking rolls out for developers in Gemini API and AI Studio; for enterprises in private preview in Gemini Enterprise (coming soon to Gemini Enterprise for Customer Experience and Google Workspace business customers); for everyone in Gemini Live, and for Google AI Pro and Ultra subscribers in Workspace Docs Live, with all Google AI subscribers in Gmail Live and Keep Live.

That is two SKUs, two rollout matrices. Buying “Live” without naming which string and which surface is how pilots stall.

What works

Fluid conversation with work in the background is the core claim. Both models can execute tools and API calls while continuing to talk, so the agent can acknowledge a request and keep chatting while tasks finish. Extended Thinking adds simultaneous reason-and-speak behavior: early verbal cues such as acknowledging the prompt, plus live progress narration through multi-step background work. Google demos (product blog) show onboarding with visual context, chess with live vision, sketch-to-React with voice feedback, multi-step bookings with async function calls, and spoken business-plan builds. Treat demos as demos.

Benchmarks, as Google stated them on September 15, 2026: Gemini 3.8 Live Extended Thinking takes the #1 overall spot on Artificial Analysis’ Speech to Speech Quality Index at 82.6, leads agentic task completion with 68.6% on τ-Voice and 35.1% on Sierra’s τ-Voice-banking benchmark, and scores 97.7% on Big Bench Audio, while Google calls the price “highly competitive” versus other frontier models. Gemini 3.8 Live is described as highly preferred among users and second place in the Speech Agent Arena, with a cost-efficient profile for scale. On ServiceNow’s EVA-Bench, Google says the models push the Pareto frontier for complex workflows by balancing accuracy with conversational quality (note: run on the Live API on Gemini Enterprise Agent Platform).

Multilingual and visual features are concrete in the posts: automatic detection and mid-conversation transitions across 97 supported languages (product blog; developer post says 97+), near real-time visual inputs, alphanumeric precision for codes and claim numbers, and incremental content updates that merge live audio with structured data.

Transparency: all audio generated by Google’s AI products is watermarked with SynthID, an imperceptible watermark in the audio output. The product blog points to the model card for safety detail.

Developer pricing, from the Fortin/Schaeff developer post: competitively priced at $0.005 per minute for audio input and $0.018 per minute for audio output (footnote estimates based on $3/1M tokens input and $12/1M tokens output). SiliconANGLE (September 15) reports the same audio meter for standard 3.8 Live, and adds that Extended Thinking also charges for reasoning tokens and for additional inputs such as video and documents. I could not independently verify a separate public Extended Thinking premium table beyond that report and Google’s shared audio meter, so treat reasoning-token billing as a confirm-on-console item before you forecast unit economics. The SKU split is still primarily capability and surface access, not a single line item.

What breaks

Access honesty first. Enterprise is private preview. Workspace business customers are “coming soon” for Extended Thinking. Docs Live is gated to Google AI Pro and Ultra subscribers. If your procurement sheet assumes GA voice agents across Gemini Enterprise for Customer Experience tomorrow, the blog does not support that.

Client state bugs will ship if you treat Extended Thinking like a chat completion API. The turnComplete semantics change is a production footgun. Teams that hang up the socket when the turn flag flips will drop tool results and mid-reason audio. Live’s migration notes also matter: proactive audio cannot be disabled, affective dialogue flags must be removed, and video frames default into context (cost and privacy implications if you stream camera when you do not need it).

Extended Thinking refuses synchronous blocking tools. If your voice stack still assumes the model waits silently while a CRM write returns, you must redesign for NON_BLOCKING behavior and narration. That is not a config tweak. It is an interaction design change.

Vendor benchmarks are not your SLA. Speech to Speech Quality Index 82.6, τ-Voice 68.6%, τ-Voice-banking 35.1%, Big Bench Audio 97.7%, Speech Agent Arena second place, EVA-Bench Pareto claims: all are Google-reported on September 15. Artificial Analysis and Sierra benches are third-party frameworks, but the scores arrive through Google’s blog. Rebuild latency, barge-in, tool success, and escalation rates on your traffic before you retire a cascaded ASR-LLM-TTS stack.

Known limitations from the model card: general foundation-model issues including hallucinations; ongoing jailbreak work; occasional slowness or timeout issues; knowledge cutoff January 2025. Voice agents that need fresh policy or price truth still need tools and retrieval, not vibes.

SynthID helps detection. It does not stop a wrong spoken refund promise. Compliance still owns disclosure, recording consent, and what the agent is allowed to say while tools run in the background.

Who it is for

Choose 3.8 Live when you need scale, cost control, and low-latency dialogue: high-volume support deflection, simple booking flows, multilingual reception, Search Live style Q&A, and agents where interleaved reasoning is enough and you do not need configurable deep thinking levels.

Choose 3.8 Live Extended Thinking when the call is the work: multi-step bookings, claims that need tool chains, sketch-to-code style collaborative sessions, Docs/Gmail/Keep Live productivity, and any workflow where the user must hear progress while the model reasons. Developers get both in API and AI Studio. Enterprises should assume private preview gates until their account team says otherwise.

Not a fit (yet) if you need public GA enterprise CX everywhere, offline-only air-gapped voice, or a single SKU that magically covers both cheapest scale and deepest live reasoning without routing logic. Also not a fit if your legal team cannot accept SynthID-watermarked AI audio in customer channels, or if your stack cannot implement async tool protocols.

Marketing, GTM, and agency operators: this is the voice-agent buy that will show up in RFP tables next to “real-time agent.” Force the two-row table. Creative and CX teams using Workspace Live features need to know which Google AI subscription unlocks Docs Live versus Gmail/Keep Live.

Pricing and limits

Developer meter (September 15 developer blog): $0.005/min audio in, $0.018/min audio out for these Live models, with the token-based footnote above. SiliconANGLE says Extended Thinking also meters reasoning tokens plus extra modalities. Confirm both the audio rate and any reasoning-token line on your project billing before you call the SKUs cost-equivalent.

Token limits (API docs): 131,072 input, 65,536 output for both model pages. Model card rounds context to up to 128K input and 64K output. Caching, code execution, file search, structured outputs, Maps grounding, image generation, URL context, and Batch API are listed as not supported on these Live model pages. Function calling: supported (Live supports async with blocking still available; Extended Thinking async only). Search grounding: supported. Thinking: interleaved on Live; configurable background thinking on Extended Thinking.

Rollout limits again, because they are the real price for enterprises: private preview in Gemini Enterprise; CX enterprise surface coming soon; Workspace business Extended Thinking coming soon; consumer and subscriber Workspace Live surfaces as listed above. Model card distribution lists Gemini API, Gemini App, AI Studio, Google Cloud / Vertex AI, Search Live for 3.8 Live, and adds Google Workspace (Gmail, Docs, Keep) for Extended Thinking.

Safety distribution note: DeepMind says Gemini 3.8 Audio did not show meaningful new frontier capabilities versus Gemini 3.7 Flash assessments under their Frontier Safety Framework framing. That is a safety process claim, not a product quality claim.

What to do this week

1. Split the purchase order. Two model IDs, two success metrics. Live: latency, containment, cost per minute. Extended Thinking: multi-step completion, tool success after IDLE, user-rated progress narration.

1. Patch the client for Extended Thinking before any pilot. Handle interaction_status, keep the session open after turnComplete, and ban sync blocking tools on that model string.

1. Map surfaces to subscriptions. Search Live, Gemini Live, Docs Live, Gmail Live, Keep Live are not one entitlement. Write which seats get which surface before demos hit executives.

1. Run a bake-off on three call types you actually own: simple FAQ, tool-heavy booking or claim, and a visual or Workspace-assisted task. Score with your own timers and escalation rates. Do not paste 82.6 into a board deck as if it were your number.

1. Decide watermark and disclosure policy for SynthID audio in customer channels. Legal and brand need a one-pager before volume ramps.

1. Keep a cascaded fallback for knowledge-cutoff and timeout cases. January 2025 cutoff plus occasional slowness means tools and human handoff remain part of the design, not an apology path.

1. Ask Google or your Cloud rep for enterprise preview scope in writing: Gemini Enterprise only, CX timing, Workspace business timing, and whether Vertex paths match AI Studio model strings.

Sharp close

Google did not ship one “Live” brain with a marketing adjective. It shipped a scale voice SKU and a high-complexity speak-and-reason SKU, with different thinking controls, different tool rules, and different doors into Search, Gemini, Workspace, and enterprise preview. Operators who buy voice agent as a single line will pick the wrong latency budget, the wrong async contract, and the wrong seat map. Live is two SKUs. Route on purpose, or your pilot will measure the model you did not actually deploy.