Gemini 3.8 Flash TTS Pricing: What 1 Minute, 1 Hour and 10,000 Minutes Really Cost

Quick answer: translate audio tokens into minutes

Google lists Gemini 3.8 Flash TTS at a price per million audio tokens. That number is accurate—and awkward if your budget is measured in narration minutes, podcast hours, or a localization backlog. Google's documented duration conversion makes those workloads calculable without guessing speech speed; the figures immediately below show the resulting token counts and audio-output costs.

Those are audio-output figures, not a promise about your final invoice. Gemini TTS also bills the text tokens you send. Audio dominates the round examples below, but a long script, repeated instructions, or many small generations can move the input side enough that it belongs in a real budget.

The current Gemini 3.8 TTS price card

Google publishes separate text-input and audio-output rates for Flash and Flash-Lite across Standard, Batch, Flex, and Priority. Batch and Flex currently carry the same price, but they remain distinct processing modes; a shared rate does not make their API behavior interchangeable. The table keeps them on separate rows for that reason.

These rates come from the official Gemini API pricing page. The model identifiers are gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. Google announced both models as generally available on September 22, 2026; the calculations here use the explicit current pricing period rather than deciding from today's date.

Why a price per million audio tokens is hard to budget

Token pricing is useful for a billing system because it gives Google one metered unit. It is less useful in a planning meeting. A product owner usually knows the number of lessons, calls, chapters, or minutes to generate—not the audio-token count those assets will create. Without the conversion rate, $9 per million audio tokens has no obvious relationship to a 42-minute episode or a monthly archive job.

TTS also has two meters. The source text is input; the synthesized speech is audio output. A clean duration comparison should isolate audio first, then add measured input tokens. Blending both into a mysterious per-minute estimate makes the result look simpler while hiding the assumption that produced it.

audio_tokens = seconds × 25

audio_output_cost = audio_tokens / 1,000,000 × audio_rate

full_generation_cost = audio_output_cost + text_input_cost

What one minute of generated speech costs

The one-minute conversion is short: seconds multiplied by Google's documented audio-token rate. At current Standard pricing, the resulting cost is a small fraction of a dollar for either model, and Batch or Flex lowers the audio portion again. This is where four-decimal pricing matters; rounding every minute to a cent would badly distort a large forecast.

A single minute is most useful as a unit rate. It lets you multiply against measured output duration and compare modes without guessing words per minute. Speaking pace may change how much text fits into that minute, but it does not change Google's documented duration-to-audio-token conversion.

What one hour costs—and why scale matters more

An hour converts to a much larger but still predictable audio-token count. The Standard cost gap between Flash and Flash-Lite is unlikely to decide a prototype by itself. It becomes material when the same hour is generated hundreds of times, localized into several languages, or repeated every month. The workload table keeps the exact, data-driven figures next to the durations.

The practical lesson is not that an hour is expensive. It is that the unit rate stays linear. Ten hours is ten times the output cost; one hundred hours is one hundred times the output cost. That makes a duration table reliable for budgeting as long as you keep input cost separate and do not add unsupported assumptions about retries, failed generations, or storage.

At 10,000 minutes, the model and mode choices stop looking trivial

Ten thousand minutes is a useful planning unit for a large content library, a podcast archive, an audiobook pipeline, localization work, or a portfolio of voice features. It is not presented as any company's measured traffic; it is simply large enough to expose differences that disappear in a one-minute example. Its exact hours and audio-token total are shown in the computed workload table above.

At current rates, the mode decision can change the output bill as much as the model decision. The computed table shows Standard, Batch, and Flex side by side for the same duration. Whether a mode fits your system is a separate operational question; price alone does not establish eligibility or acceptable processing behavior.

Flash-Lite saves one third on the same audio workload

For the same audio duration under current Standard pricing, Flash-Lite has a lower audio-output charge than Flash. The computed comparison below reports both the absolute difference and the percentage reduction for 10,000 minutes. The percentage holds for other durations because both models use the same documented conversion and the calculation is linear.

That statement is intentionally narrow. It says what the verified rate cards do to an equal audio workload. It does not say Flash-Lite sounds worse, Flash sounds better, or either model is the right production choice. Voice suitability needs listening tests with your scripts, languages, style controls, and acceptance criteria.

Batch and Flex can halve the audio-output cost

Current Batch and Flex audio rates are half their Standard counterparts for both models. The math is immediate: the same 15 million output tokens cost half as much. Priority moves in the other direction, with a higher rate than Standard. A useful cost review should therefore ask which processing mode the workload actually uses before debating small prompt optimizations.

Do not read the matching Batch and Flex prices as matching products. Google documents them as separate modes and exposes each independently. Confirm availability and processing semantics in the official documentation, then run the exact mode through the calculator. The article shows the economic range; it does not choose an operating model for you.

The January 2027 price increase doubles these examples

Google's current Gemini 3.8 TTS rates end on December 31, 2026, and the published future period starts January 1, 2027. The data-driven table below shows both Standard audio-output rates and their workload effect; text-input rates change in the same period. A forecast that crosses New Year's Day should not extend the launch-period numbers indefinitely.

For a recurring 10,000-minute monthly workload, the table reports the added monthly audio-output spend and its twelve-month impact for each model. This is a clean price-period comparison: it assumes the same duration every month and excludes text input, retries, and other services.

Your real bill also includes text input

Duration answers the biggest question, but it is not the whole request. Suppose one Standard Flash generation produces 60 minutes of audio and sends 12,000 text input tokens. The audio side remains 90,000 tokens. The input side is priced independently at the model's current text-input rate, then added to the output charge.

The example is deliberately simple. It does not claim that every hour of speech contains 12,000 input tokens, because tokenization, markup, instructions, and language differ. Use the token count from a representative API request. For a monthly plan, multiply each generation's full cost by the number of generations rather than multiplying audio alone.

Keep the estimate tied to verified pricing

Use the Google Gemini API pricing page to review source-linked provider records, the Model Price Monitor to check recorded pricing changes, and the AI API Pricing Dataset for the structured data behind the site's pricing tools. If your workload also uses ordinary text-generation models, estimate that part separately with the AI API Cost Calculator.

Provider pricing can change, and an invoice can include usage that a duration-only example cannot see. Treat these calculations as an auditable planning layer: the conversion is explicit, the pricing period is explicit, and the text-input caveat is visible. Then compare the estimate with your billing export once real traffic exists.

FAQ