Research-backed calculator guide

Gemini Live API Voice Agent Cost: Input Audio, Output Audio, and Barge-In

Estimate Gemini Live API voice-agent spend from audio seconds, session turns, context growth, transcription, tools, and interrupted output.

What the Live API meters

Gemini Live is a stateful, bidirectional API designed for continuous audio, video, and text streams. Google documents raw 16-bit PCM audio at 16 kHz for input, raw 16-bit PCM at 24 kHz for output, and a stateful WebSocket connection ^[2]. That architecture matters for cost because the bill follows the model’s input and output tokens, not the size of a telephone bill or the number of WebSocket packets.

The most important meters for a voice agent are:

1. Current audio input. Microphone audio sent to the model is converted using the published 25-token-per-second estimate. At $3 per million input audio tokens, one minute of continuously transmitted input is approximately 1,500 tokens and $0.0045, which Google displays as roughly $0.005 per minute ^{[1] [3]}.
2. Audio output. Spoken model output uses the same 25-token-per-second conversion. One minute is approximately 1,500 output audio tokens. At $12 per million, that is $0.018 per minute ^{[1] [3]}.
3. Text and reasoning. Text input, text output, and thinking tokens use their own rates. The current Gemini 3.8 Live table lists $0.75 per million text input tokens and $4.50 per million text output tokens; output pricing includes thinking tokens ^[1]. Do not multiply a text token rate by audio seconds.
4. Accumulated context. The Agent Platform pricing documentation says the session context window includes the new turn plus accumulated previous turns, and that tokens present in that window are charged on each turn ^[3]. This is the cost multiplier that a headline “per-minute” number hides.
5. Optional features. Audio-to-text transcription is charged at the standard text output rate for generated transcription tokens ^[3]. Grounding with Google Search, functions, your own server, and any telephony provider should be modeled as separate line items. Google’s Live API overview confirms tool use, audio transcription, and barge-in as capabilities, but capability support is not the same thing as a bundled price ^[2].

The exact model and billing surface still matter. The Developer API pricing page and Google Cloud’s Agent Platform table are the authoritative places to check before launch. Rates, model names, preview status, and regional SKU treatment can change, so preserve the checked date and model ID in your estimate.

Worked cost example

Consider a support agent with one ten-minute session. Assume five minutes of user speech, five minutes of model speech, six conversational turns, no video, no transcription, no grounding, no tool calls, and no telephony charge. The first calculation isolates the two audio meters.

The $0.1125 result is more precise than multiplying ten elapsed minutes by both rates, because the user and model do not speak for the entire elapsed session at the same time. If both sides were active for ten minutes each, the raw audio subtotal would be $0.225. Write down talk time separately from wall-clock session time.

Now add session context. The following is an intentionally transparent scenario estimate, not a Google quote. Assume six turns; each turn adds 50 seconds of user audio and 50 seconds of model audio, or 1,250 tokens for each side. Also assume a 1,000-token text system instruction remains in context. For a conservative worksheet, count the current user audio plus all previous user and model audio as input-context tokens on each turn, then charge the six model responses separately as output audio.

The input-context sum is:

• System instructions: 6 × 1,000 = 6,000 tokens
• Current and previous user audio across six turns: 1,250 × (1 + 2 + 3 + 4 + 5 + 6) = 26,250 tokens
• Previous model audio carried into later turns: 1,250 × (0 + 1 + 2 + 3 + 4 + 5) = 18,750 tokens
• Total modeled context input: 51,000 tokens

At $3 per million input tokens, that context component is $0.153. Six current model responses at 1,250 output tokens each total 7,500 output tokens, or $0.090 at $12 per million. Under this simplified context-replay scenario, the modeled total is $0.243 before optional features. It is higher than the raw $0.1125 audio-meter subtotal because the same conversation history is assumed to be billed again across turns.

Do not copy $0.243 into a forecast without checking your implementation. Context compression, session management, the exact turn boundaries, retained content, and model response shape change the result. The value of the worksheet is that each assumption is visible and can be replaced with usage records.

How barge-in affects the budget

Barge-in means the user can interrupt the model while it is speaking. Google lists this as a Live API feature ^[2]. Operationally, your client needs to stop playing the current audio, send the user’s new activity, and record which output chunks were actually received and played. The pricing pages do not state that an interrupted response receives a pro-rated refund or a special free treatment.

For planning, keep two numbers:

• Generated output audio: what the API returned before cancellation.
• Played output audio: what your application delivered to the user.

The first is the safer provisional cost meter; the second is a product-quality metric. Reconcile the provisional value against response usage metadata and your provider invoice. If a call has frequent interruptions, model a higher turn count and shorter average response, then compare that scenario with a low-interruption baseline. Do not quietly subtract all unplayed audio unless your usage record proves that it was not billed.

Barge-in can also raise context costs. An interrupted answer may create a new user turn while earlier model content remains in session history. A good ledger records session ID, turn ID, input audio seconds, output audio seconds, cancellation time, retained context tokens, transcription tokens, and tool or grounding calls. That row structure makes it possible to explain why two ten-minute calls with different interruption patterns do not cost the same.

What is outside the Gemini estimate

A Gemini model estimate is not a complete voice-agent cost. Add the following separately when they apply:

• Telephony or SIP. Phone numbers, carrier minutes, call recording, and media gateways are not Gemini audio tokens.
• Transport and compute. WebSocket servers, autoscaling, logging, storage, monitoring, and egress belong to your infrastructure budget.
• Speech transcription. If you enable audio-to-text transcription, generated text tokens are billed at the applicable text output rate ^[3]. Your own storage or analytics pipeline may add cost.
• Tools and grounding. Function calls can trigger CRM, booking, search, or payment systems. Grounding has its own usage rules; do not bury it inside the audio subtotal ^{[1] [3]}.
• Video and images. Sending frames changes modality and token consumption. Exclude video from a voice-only estimate and create a separate frame-frequency scenario.
• Application retries. Reconnects and duplicated turns can create real usage. Keep request IDs and usage records so retries can be distinguished from successful user turns.

For a monthly budget, multiply sessions by the distribution of session types rather than one average. A useful first model has short, normal, and long calls; a high-interruption variant; and a tool-heavy variant. If you are comparing models or providers, use the Model pricing comparison route with the same workload assumptions. For token-level reconciliation, pair it with the AI token cost calculator, then validate actual invoice rows with the AI API invoice reconciliation calculator.

A practical worksheet for production estimates

Start with observed audio seconds, not a guessed “minutes per customer” number. For each session, capture user input seconds and model output seconds separately. Convert each side to tokens with the 25-token-per-second planning factor, then apply the model’s current input and output rates. Next, add text input, text output, thinking, transcription, grounding, tool, and infrastructure rows. Finally, calculate p50 and p95 session cost. The p95 view is useful for budget headroom, while the mean is useful for unit economics.

Keep vendor facts and workload assumptions in different columns. The vendor facts here are the published rates and conversion guidance dated November 2, 2026 ^{[1] [3]}. The balanced five-minute/five-minute call, six turns, 1,000-token system instruction, and no-feature assumptions are examples created for this article. They are not statements about your traffic or Google’s average customer.

The LLM cost calculator can help you compare monthly request volume and token assumptions, but a Live API ledger should preserve audio seconds and turn IDs before aggregating. If the estimate is used for finance, retain the model ID, endpoint, region, pricing snapshot date, and an explanation of excluded services. When an invoice differs, investigate meter and context differences before changing the unit rate.

FAQ

Is Gemini Live API cheaper per minute than ordinary text API usage?

It is not a like-for-like comparison. Gemini Live publishes audio input and output rates, while ordinary text requests are measured in text tokens. A voice session can look inexpensive on raw audio seconds and still cost more than expected when accumulated session context, thinking, transcription, tools, or telephony are added. Compare the same completed task and traffic shape, not just a headline per-minute number ^{[1] [3]}.

What is the Gemini audio output price?

For Gemini 3.8 Live on the checked paid-tier table, audio output is $12 per 1 million audio tokens, displayed as approximately $0.018 per minute using 25 audio tokens per second ^[1]. That is a model-output estimate. It excludes context replay and external voice infrastructure.

How much does a ten-minute Gemini voice call cost?

Under the narrow example of five minutes of user speech and five minutes of model speech, with no context replay or optional features, the calculated audio subtotal is $0.1125, or about $0.115 using the displayed minute estimates. A six-turn session with a retained 1,000-token instruction and cumulative context modeled as billed each turn produces an illustrative $0.243 total. Use the second number only as a scenario until your usage records confirm the retained context and turn accounting ^[3].

Does barge-in reduce Gemini Live cost?

Google documents barge-in as a capability, but the cited pricing documentation does not publish a separate interruption discount ^{[1] [2]}. Track generated output, cancellation timing, retained context, and usage metadata. Budget conservatively until you can reconcile whether cancelled output was included in billed usage.

Does transcription add to Gemini Live API cost?

It can. Google’s Agent Platform pricing notes that audio-to-text transcription is charged at the standard text output rate for the generated text tokens ^[3]. Add transcription tokens as a separate line item rather than assuming they are included in the audio-output rate.

Does this estimate include a phone number or call provider?

No. The examples cover modeled Gemini usage only. Telephony, SIP, carrier minutes, media relays, recording, servers, monitoring, storage, and egress must be added from your own vendor quotes and traffic assumptions.