Google's new voice AI can talk, see and get work done at the same time

On September 15, 2026, Google launched two real-time voice models:

• Gemini 3.8 Live — "built for scale and cost efficiency," combining conversational intelligence, fluid dialogue, visual grounding (near real-time video/image input), automatic mid-conversation switching across 97 languages, and background execution of tools/API calls while the conversation continues. Priced at $0.005/minute audio input and $0.018/minute audio output. Rollout: Gemini API, Google AI Studio (developers); private preview in Gemini Enterprise (enterprises); Search Live in the Google app (consumers).
• Gemini 3.8 Live Extended Thinking — "built for high-complexity tasks," adding configurable background reasoning (thinking levels low/medium/high) that runs while the model keeps speaking ("reasons and speaks simultaneously"), asynchronous non-blocking tool calls only, and live progress narration ("Let me check that…") that eliminates conversational dead air. Google states it holds the #1 spot on Artificial Analysis' Speech-to-Speech Quality Index (82.6). Rollout: Gemini API / AI Studio; private preview Gemini Enterprise; Gemini Live app; Workspace Docs Live (Google AI Pro/Ultra subscribers); Gmail Live and Keep Live (all Google AI subscribers).

Both are available via the Gemini Live API (stateful WebSocket, raw 16kHz PCM audio in, JPEG frames ~1fps, streaming audio out), with day-one integrations from Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents, and launch partners including Salesforce, Genspark and Lumeris. All AI-generated audio is watermarked with SynthID.

FACT: Google now offers production-grade, natively speech-to-speech voice models — no ASR→LLM→TTS cascade required — that can see, reason and act inside a single live session.

Key deltas versus the previous state:

• Latency architecture: single native speech-to-speech model vs. cascaded pipelines that Google/third parties estimate add 450–1,000 ms; Google positions sub-200 ms end-to-end (COMPANY CLAIM / EARLY RESEARCH — not independently measured at this time).
• Reasoning model split: a low-latency volume SKU (3.8 Live, interleaved reasoning) and a high-reasoning SKU (Extended Thinking, background reasoning + narrated progress) — a deliberate two-SKU product architecture replacing the earlier gemini-3.1-flash-live-preview in the Live API.
• Conversational state machine: with Extended Thinking, turnComplete no longer signals the end of an interaction; clients must track interaction_status (IN_PROGRESS / IDLE). This is a new client-side contract for voice-agent developers.
• Async-by-default tools: non-blocking NON_BLOCKING function calling is the default on 3.8 Live and the only mode on Extended Thinking (blocking returns a hard error) — tools run while audio streams.
• Visual grounding in consumer surfaces: Search Live (Google app "Live" icon) lets users point the camera at something and ask questions conversationally, with transcript links and AI Mode history.
• Pricing: $0.005/min in / $0.018/min out (~$1.38/hour for overlapped conversation using published rates), undercutting OpenAI's GPT-Live-1 front-end at $0.05/min (≥$3.00/hour) — COMPANY CLAIM pricing confirmed by independent outlets (The Decoder, TPS, BitsMinds, BibiGPT), and roughly consistent with Artificial Analysis hourly cost measures reported by datanorth.ai ($0.84/hr for 3.8 Live, $3.50/hr for Extended Thinking input-audio, vs $5.83/hr GPT-Live-1 Astra).

Before (Sep 10–14, 2026): OpenAI's GPT-Live-1 (Sep 10) set the frontier for real-time voice with full-duplex conversation at $0.05/min (backend billed separately); Google's best live option was gemini-3.1-flash-live-preview and Gemini 3.8 Flash (Sep 2) — a multimodal speed model, not a dialogue model; most voice agents still ran cascaded ASR+LLM+TTS stacks that stalled ("dead air") during tool calls; voice assistants were largely turn-based.

Change (Sep 15, 2026): Two stable native speech-to-speech model IDs (gemini-3.8-live, gemini-3.8-live-extended-thinking) in the Live API with in-session reasoning, visual grounding, 97-language auto-switching, background async tools, narrated progress, and a ~7x cheaper hourly price point than the OpenAI front-end (on Artificial Analysis' input-audio measure).

After: Developers can build voice agents that talk, see, reason and act concurrently; Google's default consumer surfaces (Search Live, Gemini app, Workspace Docs/Gmail/Keep) run on the new models; enterprises can private-preview the models in Gemini Enterprise; the industry's voice-agent default UX shifts from "silence while thinking" to "narrated background work."

FACT (from official docs and model card): Both models are natively multimodal audio models within the Gemini 3 series ("Gemini 3.8 Audio"), processing continuous streams of audio, video and text to deliver spoken responses in real time. Technical properties:

• Inputs/outputs: text, images, audio, video in; text + audio out. 131,072-token input window (model card: "up to 128K"), 65,536-token output.
• Transport: Gemini Live API over a stateful WebSocket; raw 16 kHz PCM audio in, JPEG images (up to ~1 frame/second), text; ~24 kHz audio out (per third-party documentation summaries).
• 3.8 Live: interleaved reasoning optimized for latency; NON_BLOCKING async function calling is default (blocking mode available for backward compatibility); full-session send_client_content with explicit user/model roles; video frames included in turns by default.
• 3.8 Live Extended Thinking: background reasoning controlled by thinking_config (thinking_level: low/medium/high; MINIMAL not supported); async non-blocking tools only (function scheduling not supported); proactive audio permanently enabled; clients monitor interaction_status (IN_PROGRESS → IDLE) because turnComplete only ends an utterance; the model can emit multiple spoken segments (acknowledgements, progress narration, final answer) within one interaction.
• Safety: every AI-generated audio output is watermarked with SynthID; model card notes the Frontier Safety Framework assessment (based on Gemini 3.7 Flash results) finds no Tracked/Critical Capability Levels likely; knowledge cutoff January 2025.

COMPANY CLAIM: "sub-200 ms" end-to-end latency positioning and the specific benchmark scores below.

Benchmarks (COMPANY CLAIM unless noted): Extended Thinking — Artificial Analysis Speech-to-Speech Quality Index 82.6 (#1 overall), τ-Voice agentic task completion 68.6%, Sierra τ-Voice-banking 35.1%, Big Bench Audio 97.7%; 3.8 Live — 2nd place in Speech Agent Arena; ServiceNow EVA-Bench Pareto-frontier results for complex workflows. Independent measurement (Artificial Analysis, as reported by datanorth.ai): Extended Thinking 82.6 vs GPT-Live-1 Astra 81.5; τ-Voice 68.6% vs 67.9%; τ-Voice-banking 35.1% vs 32.0%; Full Duplex Bench 91.9% vs 94.9% (OpenAI still leads); time-to-first-audio 1.35 s vs 1.34 s; hourly input-audio cost $3.50 vs $5.83 (base 3.8 Live: ~$0.84/hr, 76.0 index score). Note: datanorth.ai flags the GPT-Live-1 row as a single trial (most models get three), i.e., the OpenAI figure is noisier.

1. The voice-interface race just got a price anchor. Google has undercut OpenAI's GPT-Live-1 (and xAI's Grok Voice Think Fast 2.0) on cost by a wide margin while roughly matching or edging ahead on quality indices — precisely the "scale + cost" play Google has run successfully before.
2. "Dead air" is now a solved-looking problem. Extended Thinking's narrated background reasoning changes user expectations for every voice agent; competitors' UX will be judged against it.
3. Distribution. This upgrades the default assistant experience across Google's surfaces (Search, Android, Workspace, Gemini app) — billions of users — turning voice-agent capability into a default feature, consistent with the discovery's "billions of Android/Google users" framing.
4. Architectural statement. Keeping speech, reasoning and tool execution in one stateful session (vs. OpenAI's split front-end/backend design) is a bet that simplicity wins developer adoption.
5. Week's context: lands five days after OpenAI's GPT-Live-1, two days after Apple's Siri AI event (Sep 14), in the same week as xAI's Voice Agent API ($0.08/min, Sep 17) — the voice-agent platform war is now the week's most concentrated competitive story.

• Voice agents that do real work while talking: booking flows, account investigations, refunds, troubleshooting — tools run async while the model narrates ("Let me check that…"), so complex tasks no longer require silence or "hold music."
• Vision-in-the-loop conversation: pointing a phone camera at a problem (Search Live demo: a leaky pipe) and holding a natural conversation about what the camera sees, in near real time.
• Mixed-language conversations on one model: automatic detection + switching across 97 languages mid-conversation with accent consistency.
• Production voice agents at previously impossible unit economics: e.g., ~$13,800/month for 10,000 hours of voice traffic vs $30,000+ on GPT-Live-1 front-end pricing (TPS estimate).
• In-session reasoning without backend orchestration: developers no longer need to build sideband context-passing to a separate reasoning model for every hard question.

1. Does 3.8 Live's "fluid dialogue, interruptions" match GPT-Live-1's full-duplex behavior in head-to-head use (Full Duplex Bench says no, so far: 91.9 vs 94.9)?
2. What does a real-world latency distribution look like under load (the sub-200 ms claim)?
3. When does the Live API leave Preview, and what do GA prices look like (Gemini 3.8 Flash pricing reportedly ~doubles in 2027 per Gigazine via tech-insider.org — EARLY RESEARCH)?
4. How does OpenAI respond on price, and how quickly?
5. Will Gemini Enterprise GA include agent-platform parity (EVA-Bench was run on the Live API + Gemini Enterprise Agent Platform)?
6. How do session limits evolve for video-in conversations?
7. Does background-reasoning billing (reasoning tokens + video inputs) meaningfully erode the headline price advantage?

• Near term: OpenAI answers on price or a cheaper GPT-Live tier (PREDICTION); xAI extends its agent-platform lane; independent latency/duplex evaluations of 3.8 Live appear; Live API GA with final pricing; Gemini Enterprise CX + Workspace business rollout.
• Mid term: voice-agent stacks standardize on "narrated background work" UX; Google extends Search Live-style camera conversations across surfaces; Multi-modal voice + agent platforms converge into the enterprise "voice operator" category.
• Watch items: session-limit changes for video; Gemini 3.8 Flash 2027 price increases signalling the current Live pricing is land-grab territory; the duplex-quality debate resolving via head-to-head arena results.

Google didn't win this week's voice race on naturalness — the independent duplex data, and several reviewers, say OpenAI still leads on conversational feel. Google won on the two things it has always weaponized: price (roughly 7x cheaper per hour) and distribution (Search Live, the Gemini app, Workspace, Android). The deeper story is architectural and systemic: with Extended Thinking, Google turned the voice agent into a stateful worker that talks while it works, and forced every developer to re-learn what "the conversation" means. In a week dominated by agent-misalignment alarms, the most important question in any of these launches is the one Google's model card answers with a page of safety attestations: the same background-tool autonomy that makes a voice agent feel alive is exactly the capability labs are still learning to contain. Speed, price and safety are now racing each other — and this launch put Google firmly in the lead on the first two.

• FACT: announcement date/rollout, model IDs, pricing figures as published, context/session limits as documented, partner list, SynthID watermarking, API semantics (async-only on Extended Thinking, interaction_status), knowledge cutoff.
• COMPANY CLAIM: "most advanced live dialogue models," benchmark scores Google quoted (82.6 #1, τ-Voice 68.6%, τ-Voice-banking 35.1%, Big Bench Audio 97.7%, Speech Agent Arena 2nd, EVA-Bench Pareto), "sub-200 ms" latency, 97-language mid-conversation switching robustness, "built for scale and cost efficiency."
• INDEPENDENT EVIDENCE: Artificial Analysis measurements as reported by datanorth.ai (82.6 vs 81.5; 91.9 vs 94.9 Full Duplex; 1.35 s vs 1.34 s TTF; $0.84/$3.50 vs $5.83 per-hour), same-day confirmation of date/rollout by multiple outlets, price comparisons in The Decoder/TPS/BitsMinds/BibiGPT.
• INTERPRETATION: price-vs-quality tradeoff reading (The Decoder/TPS), in-session-vs-delegated architecture split (The New Stack/BibiGPT), cost leadership playbook, "voice as next OS surface."
• PREDICTION: OpenAI price response; Live API GA timeline; voice-agent UX standardization on narrated background work.