Gemini 3.8 Live API Pricing 2026: Voice Agents

Google DeepMind's First Audio-to-Audio Lineup in Six Months

On September 15, 2026, Google DeepMind published Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking, the company's first pair of native audio-to-audio models since Gemini 2.5 Flash Native Audio shipped in December 2025. The two endpoints slot into the Live API at gemini-3.8-live and gemini-3.8-live-extended-thinking, and they sit at the top of every major speech-to-speech leaderboard that Artificial Analysis publishes: 3.8 Live Extended Thinking (High) at 82.6 on the Speech-to-Speech Quality Index, 3.8 Live at 76.0. For an API team building a voice agent, this is the first model pair in 2026 where Google's native audio model is genuinely competitive with — and in most measurements ahead of — GPT-Realtime, Grok Voice, and the open-weights audio stack.

The pricing is also genuinely new. The 3.8 Live endpoint charges $0.75 per million text input tokens, $4.50 per million text output, $3.00 per million audio input ($0.005 per minute of audio in), and $12.00 per million audio output ($0.018 per minute of audio out). That is meaningfully cheaper than GPT-Realtime-2 (High) on a per-hour-of-input-audio basis ($0.84 versus $4.14 according to Artificial Analysis), and the Extended Thinking variant sits at $3.50/hour — comparable to GPT-Live-1 (Astra, medium) and below GPT-Realtime-2 (High). What is genuinely new is the dual billing: text and audio are priced independently within the same Live session, so a voice agent that interleaves typed system prompts with spoken user turns pays the text rate for the typed half and the audio rate for the spoken half.

Two Endpoints, One Pricing Table

Google ships 3.8 Live as the default Live API model and 3.8 Live Extended Thinking as the reasoning-heavy variant. The DeepMind announcement explicitly recommends updating from gemini-3.1-flash-live-preview to gemini-3.8-live, and the 3.1 preview is deprecated within weeks. The two endpoints share a single pricing table on Google's API pricing page because the work they do is structurally similar — they are both audio-to-audio models that stream speech in and speech out — but they differ on three axes that an API team will care about.

• Reasoning depth. 3.8 Live runs interleaved reasoning inline with audio output. 3.8 Live Extended Thinking runs background reasoning as a parallel task and emits verbal cues like "Let me check that..." while the reasoning continues. For a call-center agent where most turns are short and the answer is in the script, the standard model is enough. For an agent that has to look something up, run a tool call, or reason through an ambiguity mid-sentence, Extended Thinking is the right choice — and the DeepMind announcement specifically demos it on employee onboarding, multi-step bookings, business-plan composition, and React component generation from sketches.
• Latency. 3.8 Live standard hits 1.18s time-to-first-audio (Artificial Analysis, September 15). Extended Thinking (High) is 1.35s — slower because the reasoning has to start before the first audio frame streams. Both are faster than the previous-generation 3.1 Flash Live High at 2.99s, which is the reason Google deprecates the 3.1 preview.
• Token burn. Standard 3.8 Live costs $0.84 per hour of input audio on the Artificial Analysis methodology. Extended Thinking (High) costs $3.50 per hour — about 4× per hour. The unit prices are identical; the difference is volume, because Extended Thinking reasons and speaks simultaneously and emits more output tokens per turn.

The Verified Pricing: Audio, Text, and Video Are Independently Priced

Verified against ai.google.dev/gemini-api/docs/pricing on September 15, 2026. The 3.8 Live and 3.8 Live Extended Thinking group shares a single Standard and Batch table; the prices below are what the Live API actually bills per modality.

Three pricing facts that matter for an API team and that the DeepMind announcement does not spell out.

• Output tokens are 6× input tokens for text and 4× for audio. A live agent that speaks more than it listens will be output-bound, not input-bound. For a support agent, plan your cost model around audio output minutes, not audio input minutes.
• Thinking tokens are billed at the output rate. Extended Thinking is not free at the margin — every reasoning step it runs is part of the output token count. This is why the $3.50/hour figure for Extended Thinking is so much higher than the $0.84/hour for standard 3.8 Live: same unit price, four times the output tokens per minute of speech.
• The Free Tier has an "improve our products" opt-in. Free-tier inputs and outputs may be logged to improve future models. This disqualifies the Free Tier for any production workload that touches PII, regulated data (HIPAA, GDPR-controlled), or proprietary code. For a personal hobby project the Free Tier is fine.

Batch Mode Is Half Off for Text, But Audio Has No Batch

Google publishes a separate Batch pricing row for the 3.8 Live group: $0.10 per million text input tokens, $0.225 per million image input ($0.00006 per image), $3.25 per million audio input ($0.00008 per second), and $6.00 per million video input ($0.000395 per frame). Batch on the Live API works the way Batch works on the rest of the Gemini API: you upload a JSONL of requests to ai.google.dev/batch or call the Batch endpoint, get the result back asynchronously within 24 hours, and pay half price.

The catch is that Batch does not preserve the streaming audio-to-audio nature of the Live API. Batch on a Live endpoint means you bundle offline speech-to-speech jobs — pre-recorded audio to transcript, or batch translation jobs — and get back a JSON with the result. It does not mean you can run a real-time voice agent at half price. For a real-time voice agent the only billing that matters is the Standard tier above.

Benchmarks: 82.6 on STS-QI, 68.6% on τ-Voice

The DeepMind announcement, the per-model ai.google.dev/gemini-api/docs/models/gemini-3.8-live page, and Artificial Analysis' public speech-to-speech index all converge on the same picture: 3.8 Live Extended Thinking (High) is the highest-scoring audio-to-audio model published to date, and standard 3.8 Live is the cheapest audio-to-audio model at any reasonable quality bar.

Speech-to-Speech Quality Index (Artificial Analysis, Sept 15, 2026)

Agentic and Audio Benchmarks

• τ-Voice (overall): Gemini 3.8 Live Extended Thinking (High) at 68.6% leads, ahead of GPT-Live-1 (Astra, medium) at 67.9% and GPT-Live-1 (Sol, low) at 59.3%. Standard Gemini 3.8 Live at 30.1% — Extended Thinking is the only variant that competes on the agentic slice.
• τ-Voice-airline: Extended Thinking at 72.0% vs. 3.1 Flash Live High at 54.7%. The biggest gap on the airline slice, where the agent has to follow multi-step booking rules.
• Sierra τ-Voice-banking: Extended Thinking at 35.1% per the DeepMind announcement — a benchmark Google cites for voice-agent task completion on regulated financial workflows.
• Big Bench Audio (BBA): Extended Thinking at 97.7%, Step-Audio R1.1 at 97.6%, standard 3.8 Live at 91.7%. The BBA leaderboard measures how well a model handles interruptions, background noise, and mid-utterance context switches.
• Full Duplex Bench: Standard 3.8 Live at 96.1%, the highest published score on this benchmark. The DeepMind demos (chess, React-from-sketches, multi-step bookings, business plans) lean on Full Duplex behavior.
• ServiceNow EVA-Bench: Google's models "push the Pareto Frontier for complex workflows" — meaning Extended Thinking scores higher accuracy than any prior model at the same conversational quality, and higher conversational quality than any prior model at the same accuracy. Run on the Live API on the Gemini Enterprise Agent Platform.

97 Languages, Mid-Conversation Switching, and Visual Context

The DeepMind announcement highlights three capabilities that are easy to miss because they look like marketing copy but are concrete Live API behaviors.

• 97 supported languages with mid-conversation switching. The model "automatically detects and transitions between 97 supported languages mid-conversation." For a voice agent deployed in a multilingual market this is the difference between routing each user to a language-specific endpoint and running a single model that handles every caller.
• Real-time visual context on the Live API. 3.8 Live "processes visual inputs in near real-time, enriching conversations with context for more helpful responses." The DeepMind demos include guiding employee onboarding using visual context to answer live questions, and Extended Thinking transforming raw sketches plus near real-time voice feedback into functional React components. On the Live API this is exposed as a WebRTC or WebSocket video stream alongside the audio stream — not a separate request.
• Background tool calls and function execution. The model "executes tools and API calls in the background while continuing the conversation." For an API team, this means a single Live session can hold a multi-step booking flow, a code execution, or a database lookup open in the background while the audio stream continues. The Extended Thinking variant is specifically designed for this: it runs the reasoning and the speech simultaneously.

How You Actually Call It: GenAI SDK, Raw WebSocket, ADK

Google documents three integration paths into the Live API. All three reach the same 3.8 Live endpoints with the same pricing.

• GenAI SDK (Python or Node). The recommended starting point. from google import genai; client = genai.Client(); async with client.aio.live.connect(model='gemini-3.8-live', config=...) as session. Handles WebSocket lifecycle, audio resampling (the Live API always uses 16-bit little-endian PCM, 16kHz input / 24kHz output), and tool-call dispatch automatically.
• Raw WebSocket with ephemeral tokens. For JavaScript frontends where you want the browser to maintain the WebSocket directly. Google issues ephemeral tokens scoped to a Live session; the browser opens a WebSocket to wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1beta.GenerativeService.BidiGenerateContent with the ephemeral token in the protocol header. Use this if you need lower latency than the SDK provides or if you want to avoid proxying audio through a backend.
• Agent Development Kit (ADK) Streaming. Google's higher-level orchestration layer for production voice agents. ADK handles session lifecycle, tool registration, conversation memory, and per-user state. The Live API integration is one line: Agent(...).live(model='gemini-3.8-live-extended-thinking'). Use this if you are already on ADK for other agents.

Third-party real-time-media platforms (Agora, Fishjam, LiveKit, Pipecat, Vercel AI SDK, Vision Agents, Voximplant, plus Firebase AI Logic for mobile) all integrate with the Live API over WebRTC or WebSocket. These are plumbing — they handle the audio/video transport and forward events to the Gemini endpoint. Pricing is unchanged: you pay Google for the Live API usage and the third-party platform separately.

Salesforce, Genspark, Lumeris, and the Developer Platforms

The DeepMind announcement explicitly names four categories of launch partners. This matters because it tells an API team which integrations are GA on day one and which are roadmap.

• Developer platforms (build on the Live API). Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, Vision Agents. These expose the Live API behind their own SDKs and pricing. Pick the one whose audio/video transport you already use.
• Enterprise integrators (run 3.8 Live in production). Salesforce (for Agentforce), Genspark (agent platform), Lumeris (healthcare agent platform). All three are cited as impressed with "latency, fluidity, and tool-calling capabilities."
• Google Workspace integration. 3.8 Live Extended Thinking is rolling out in Google Workspace starting September 15, with Docs Live, Gmail Live, and Keep Live. The Workspace SKU is bundled into the existing Google AI Pro and Ultra plans, not billed per-token at the Live API rate.
• Enterprise Agent Platform. The ServiceNow EVA-Bench numbers were "run on the Live API on Gemini Enterprise Agent Platform" — meaning the production path for regulated enterprise use is the Agent Platform tier of Gemini Enterprise, not the public Gemini API.

SynthID Watermarking on Every Audio Frame

Google watermarks every audio frame generated by 3.8 Live with SynthID. The announcement says it is "imperceptible" and "woven directly into the audio output." For an API team, this is a feature, not a bug: it means a voice agent built on 3.8 Live produces audibly indistinguishable output from non-watermarked speech but is detectable as AI-generated by a SynthID-aware verifier. If you are building a voice agent for a use case where provenance matters (call recording compliance, deepfake defense, agentic disclosure), SynthID on by default is a real architectural advantage over open-weights audio models. It also means you cannot strip the watermark from the audio stream the model produces without re-encoding — which is the point.

The Real Cost: A 10-Minute Support Call, Worked Out

For an API team, the practical question is what one minute of voice agent actually costs at production rates. The numbers below assume Paid Tier, no grounding, no batch.

The arithmetic behind row 4: 1,000 calls × 5 min/call = 5,000 minutes total session time, split roughly 50/50 between input and output audio. At $0.005 per minute input + $0.018 per minute output, the cost is 2,500 × $0.005 + 2,500 × $0.018 = $12.50 + $45.00 = $57.50 per 1,000 calls. The row above rounds up to include a typical 4-second spoken greeting and a 10-second spoken closing per call, plus the system-prompt text input on every turn, which pushes the per-1,000-call figure into the four-digit range. The Extended Thinking variant is 4× because every agent turn emits roughly 4× more output tokens per minute of speech — exactly the per-hour ratio Artificial Analysis reports.

The optimization lever is output audio minutes, not input. The fastest way to halve your Live API bill is to make the agent shorter — tighter system prompts, fewer spoken confirmations, no verbose greetings. The second-fastest is to handle simple calls (greetings, transfers, "I don't know" routing) on a non-reasoning routing layer and only spin up Extended Thinking when the call needs it.

When to Use Which: Routing by Call Type

For a production voice agent the right architecture is to start on the standard 3.8 Live endpoint and only escalate to Extended Thinking when the call demands it. The DeepMind announcement's own demos are a useful taxonomy.

What the Announcement Does Not Tell You

Three operational details the DeepMind post leaves out and an API team will hit in the first week.

• Session length limit. The Live API docs do not publish a hard session length cap. Long-running sessions (multi-hour onboarding calls, overnight agent loops) are likely capped by connection timeout rather than by token quota. Plan to reconnect and resume state in your application layer if your session is multi-hour.
• Concurrency limits. Google's per-project rate limit page publishes QPM and TPM but not concurrent WebSocket session caps. For a contact-center deployment with 100 concurrent agents, validate against your project quota before going to production — request a quota increase through Google Cloud Support if 3.8 Live defaults are too low.
• Per-token context window. Live API sessions are audio-to-audio, but the system prompt and any text context you push through the API are billed at the text token rate with the standard context window. There is no published "context window" for the audio stream itself — the model handles streaming context implicitly.

Migrating from gemini-3.1-flash-live-preview

Google explicitly recommends the migration. Two things change.

• Model ID string. gemini-3.1-flash-live-preview → gemini-3.8-live (standard) or gemini-3.8-live-extended-thinking (reasoning).
• Audio format. Unchanged. The Live API always uses 16-bit little-endian PCM, 16kHz input / 24kHz output. If your old code resampled to those rates, it works as-is.
• Tool/function definitions. Unchanged shape. The Live API tool dispatch is the same GenAI SDK tools= schema as the rest of the Gemini API.
• Client state management. Updated to handle asynchronous reasoning signals if you switch to Extended Thinking. The standard endpoint does not require this.

Cost changes are minor: $0.75/M text input (same as 3.7 Flash), $4.50/M text output (vs $3.75 on Flash), $3.00/M audio input (same), $12.00/M audio output (vs $12 on Flash Native Audio). The only meaningful delta is text output, which is higher because 3.8 Live emits more output tokens per turn than Flash Native Audio did — the implicit trade for the higher reasoning quality.

Should You Build on 3.8 Live Today?

If you are starting a new voice agent from scratch, the answer is yes, and the right choice is the standard gemini-3.8-live endpoint with escalation to gemini-3.8-live-extended-thinking only on multi-step reasoning calls. The pricing is the cheapest published audio-to-audio rate at any quality bar above GPT-Realtime-2, the latency is the fastest in the published set, and the Live API integration paths (GenAI SDK, raw WebSocket, ADK) are stable and documented. If you already have a production voice agent on 3.1 Flash Live or 2.5 Flash Native Audio, the migration is a model-ID swap with one pricing gotcha (text output went up modestly) and one feature gain (Extended Thinking when you need it).

If you are evaluating against GPT-Realtime-2 (High), the published comparison is: 3.8 Live standard is faster (1.18s vs 1.14s — within margin) and 5× cheaper per hour of input audio ($0.84 vs $4.14), but Extended Thinking (High) is comparable on cost ($3.50 vs $4.14) and slower (1.35s vs 1.14s). The Extended Thinking bet is on reasoning quality and tool-call reliability, not raw latency. If your agent does heavy reasoning, the bet pays; if it does not, the standard endpoint is the right answer.

Sources

• DeepMind blog, Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking — Tom Ouyang & Malini Jaganathan on behalf of the Gemini Audio Team, September 15, 2026. deepmind.google
• Google AI for Developers, Gemini 3.8 Live model reference. ai.google.dev
• Google AI for Developers, Gemini 3.8 Live Extended Thinking model reference. ai.google.dev
• Google AI for Developers, Gemini API pricing — Standard and Batch pricing tables for the Live API group, verified September 15, 2026. ai.google.dev
• Artificial Analysis, Speech-to-Speech Quality Index — public leaderboard including time-to-first-audio, cost-per-hour-of-input-audio, and τ-Voice scores. artificialanalysis.ai
• apirank, Gemini 3.8 Flash API 2026 (September 6, 2026) — the text-mode sibling of the Live API for context on the 3.8 family pricing pattern. apirank.vip