Gemini Live API Pricing: What Voice Agents Cost at Scale

“gemini api pricing” carries 4,500 monthly searches in the United States, against 8,100 for the broader “gemini pricing”. The gap is the tell. People have stopped asking what Gemini costs in general and started asking what the API costs in metered terms, which is exactly the question that gets expensive when the workload stops being text.

On September 15, 2026, Google made that transition concrete. It shipped two audio-to-audio models into general availability: Gemini 3.8 Live (gemini-3.8-live) for low-latency voice agents, and Gemini 3.8 Live Extended Thinking (gemini-3.8-live-extended-thinking) for cases where the model needs to reason in the background while the conversation continues. Google also moved the Live API from preview rate limits to general availability limits, which is the part that changes your bill.

The published per-minute rate is not the problem. The problem is which unit you get billed in, and it is not minutes.

What Gemini Live API pricing actually charges

The published rates for gemini-3.8-live are these:

• Audio input: $3.00 per million tokens, or approximately $0.005 per minute
• Image and video input: $1.00 per million tokens, or approximately $0.002 per minute
• Text input: $0.75 per million tokens
• Audio output: $12.00 per million tokens, or approximately $0.018 per minute
• Text output, including thinking tokens: $4.50 per million tokens

The minute figures require a conversion assumption, and Google made it explicit on a neighbouring model page: audio tokens are calculated at 25 tokens per second. That is 1,500 audio tokens per minute in each direction.

Run the arithmetic on a clean minute of speech. Thirty seconds of user audio at 25 tokens per second is 750 input tokens, which costs $0.00225. Thirty seconds of model speech is 750 output tokens at $12.00 per million, which costs $0.009. Call it $0.011 for a one-minute exchange with no thinking. At that blended rate a ceiling of $0.018 per minute for audio output is generous but not absurd.

Now put a real agent on the line.

The minutes lie: what a voice agent actually burns

When you look at gemini pricing pages you see per-minute rates; when the model thinks, you get billed thinking tokens at the text output rate of $4.50 per million. Three things blow past the per-minute estimate.

Reasoning on voice turns. Gemini 3.8 Live Extended Thinking exists precisely because the plain Live model will not do heavy reasoning mid-conversation. Extended Thinking runs background reasoning during a live audio session, billed as text output tokens at $4.50 per million. That number is roughly 65% of a dollar, per million tokens, on top of the audio. Google reported that the Extended Thinking variant leads agentic task completion on Sierra’s tau-Voice-banking benchmark at 35.1%, which is the kind of task that requires reasoning, not just transcription.

Video frames. The moment your agent watches anything, you are feeding image and video tokens at $1.00 per million, or about $0.002 per minute, alongside the audio stream. A persistent visual channel doubles the input legs of the meter.

Tool calls and prompt replay. The Live API supports asynchronous function calling so the model can keep talking while a tool call resolves. Every call re-sends context. Every resumed session re-sends history. None of that is unusual; all of it shows up as input tokens stacked on top of a rate you were reading as per-minute.

A minute of a working voice agent, with reasoning and a couple of tool calls, is a 30 to 90 cent minute in practice, not a two-cent minute. That is the number to build a business on.

The gemini api pricing calculator that gets it wrong

The calculator sites ranking for gemini api pricing have the same structural bug. They ask for calls per month and average call length in minutes, multiply, and print a reassuring number. That input set cannot express reasoning tokens, visual frame rate, tool-call fanout, or replayed context, which are the four levers that actually set the bill. Two voice agents with identical call volumes can differ by an order of magnitude in spend and the calculator will show them the same total.

Measure the four input legs instead. Tokens per call by modality (audio input, audio output, text, image and video), average tool calls per call, average tokens per tool round trip, and the reasoning multiplier per effort level. Log those from your own long-lived sessions for a week and you will have a model of your cost that survives a model swap, because it is expressed in tokens and not in minutes.

Track thinking separately. Thinking tokens bill at the text output rate. If you are running Extended Thinking, split thinking tokens from answer tokens in your telemetry or you will misattribute half your bill.

General availability changed the limits, not the price

Google did not change the published rates on September 15. It changed the access shape by moving the Live models out of preview. The structural relief is on the free tier: your Live traffic now falls under one shared monthly allowance of 5,000 Google Search grounding requests across all Gemini 3.x models, then $14 per 1,000 requests. Grounding is the quiet cost line on a voice agent that answers factual questions, and on preview limits it was easier to hit a wall than to hit a budget.

The paid tiers are rated by spend per rolling 10-minute window, not by requests:

• Free: no spend limit listed
• Tier 1: $10 per 10 minutes
• Tier 2: $50 per 10 minutes
• Tier 3: $200 per 10 minutes

That is the mechanism to internalize. At Tier 1, that $10 window supports roughly 4 million input audio tokens at the published rate, or about 14 cents per second of allowance. A voice agent in front of real users can burn that. The ceiling scales with your tier and with cumulative Google Cloud spend, and hitting it returns a 429 RESOURCE_EXHAUSTED error, which on a voice surface is a dropped call rather than a retry. Voice has no patience for a backoff loop.

Note also that these are ceilings, not promises: Google states plainly that specified rate limits are not guaranteed and actual capacity may vary. For a heavy user, a rate limit is a capacity statement that can move underneath you.

The Gemini 3.8 Flash trap: half price until December 31

Here is the detail that should be in every gemini api pricing calculator and is in almost none of them.

gemini-3.8-flash is priced at $0.75 per million input tokens and $3.75 per million output tokens as an introductory price that expires on December 31, 2026. On January 1, 2027, the same model costs $1.50 and $7.50, doubling overnight. Context caching follows the same curve: $0.075 per million through the end of 2026, then $0.15, with cache storage moving from $0.50 to $1.00 per million tokens per hour.

Gemini 3.7 Flash, the previous generation, carries the identical intro pricing and the identical cliff. So does the Batch tier: $0.375 in, $1.875 out through 2026, then $0.75 and $3.75.

For comparison, gemini-3.1-pro-preview is priced at $2.00 per million input and $12.00 per million output, with Priority inference at $3.60 and $21.60. And a Refresh, Flex or Batch request does not save you money on Pro: Flex is priced at the Batch rate only if your latency tolerance allows it.

What this means in practice. Any pipeline you build on Gemini Flash this quarter is being priced with a 50% discount whose expiry date is public. In a tiered architecture where a Flash model handles the long-horizon coding and agentic work and a Pro model handles the hard cases, your blended unit economics are guaranteed to worsen by roughly 60 to 70% on the Flash leg in January unless you have already re-based the workflow. Model that now, in Q4 budget terms, while you still have a quarter to move work to lower effort levels, batch it, or cache it.

The cross-provider reality: voice is the expensive modality everywhere

Gemini 3.8 Live at $12.00 per million audio output tokens sits in the same band as the rest of the market. OpenAI’s realtime audio surface prices audio output at a comparable level, and Anthropic has not shipped a realtime audio-to-audio API at all, which is why Anthropic’s voice story runs through third-party stacks rather than a first-party endpoint. Google is also framing the go-to-market the same way its competitors do: Gemini 3.8 Live rolls out to Google AI Pro and Ultra subscribers through the Gemini app and Workspace, while developers get the metered API.

That side-by-side is the interesting part for a heavy AI user. Google AI Pro is $19.99 per month. Google AI Ultra starts at $99.99 per month, with a $199.99 tier offering 20x the usage limits of Pro. At the API’s blended voice rate, the $19.99 Pro seat is roughly equivalent to nineteen dollars of metered voice output, which is about 1,050 minutes of pure audio output before any reasoning, video, or tool calls. The $99.99 Ultra tier buys roughly 5,500 metered minutes on the same arithmetic.

For a solo builder, that arithmetic should decide the architecture. If your usage fits inside a subscription surface, the subscription is cheaper than the API by design, because Google is buying your attention across Gemini, Search, Workspace, and Jules. The API is for what a subscription cannot host: your own product, your own integration points, your own multi-tenant load. Pay metered rates only for the traffic a seat cannot carry.

And note the second published constraint on the subscription side. Google documents the Gemini app’s usage model as compute-based limits that refresh every 5 hours until you reach a weekly limit, with AI credits available to extend them. That is the same three-meter structure Anthropic and OpenAI use on their top consumer plans. The pattern is now universal across Anthropic, OpenAI, Google, and as of this week Meta, whose new Meta One tiers publish no AI allowance at all. A subscription buys priority in a queue, and the metering underneath it is compute, not requests.

Build the meter before you build the agent

Log tokens per call, by modality, from session one. Not minutes. Audio input, audio output, text, image and video, plus tool-call count. A minute-based estimate is unfalsifiable until the invoice arrives.

Instrument thinking tokens separately. They bill at the text output rate and they are the single largest source of estimation error on an Extended Thinking agent.

Model the January 1 step explicitly. If you are on Flash, put the doubled rate into your unit economics now and treat Q4 as a window to move from high to lower effort levels, add context caching, or batch what can tolerate latency.

Buffer against the 10-minute spend window. Run a graduated fallback chain by effort level, and pre-provision a tier with headroom. A 429 on a voice surface is a dropped call, not a slow page.

Price the subscription alternative before committing to metered voice. At Gemini’s blended rates, a seat is roughly nineteen metered minutes per dollar on Pro and a comparable ratio on Ultra. If your volume fits a seat, the seat wins.

Add a cache-miss model to the calculator. The calculators ranking for gemini api pricing do not include one, and every long-running voice agent has a miss rate that decides its real cost.

The uncomfortable summary is the one worth keeping: Google just made a frontier voice stack generally available at rates that are genuinely competitive, and then published an expiry date on half of them. Two things are simultaneously true. Voice agents at $0.005 per minute are cheap enough to be everywhere, and the same workloads at $0.30 per minute are a business model that only closes if the agent completes real work. The published rate tells you which of those two you are running. The token log tells you the truth.