How much does it cost to make a video with AI, and how to measure it

Making a video with AI costs the sum of several items that are paid for by use (the script text, the voice, the images and the editing) and of two that appear on no invoice: your time and the number of times something has to be redone. There is no single figure, because it depends on the length, the number of scenes, the models you choose and how much you redo. What is useful is to know how each item is charged and to record the cost of each video so you can compare it with what it delivers.

The items in a video made with AI

A video narrated with AI goes through several phases, and each one has its own form of cost:

• Text: research, brief, outline, script and review. Language models are charged by input and output tokens.
• Voice: AI narration is charged by characters or by audio tokens, depending on the model. If you record it yourself, there is no AI cost, but there is a time cost.
• Images: each generated image is charged separately, and the price changes with the model and the resolution.
• Editing: combining voice, images, movements and subtitles into a video. On your computer it costs time; in the cloud, compute time.
• Your time: choosing the idea, reviewing the script, checking facts, approving scenes and preparing the publication.
• Redos: every time you regenerate a voice, an image or an edit, that part is paid for again.

How each item is charged: examples of official rates

To show how this turns into figures, these are examples from Google Cloud's official pricing pages, consulted on 9 October 2026. The prices on those pages are in US dollars and may change: check them on the official page before doing your sums.

• Text, on Vertex AI (its pricing page presents it as Agent Platform). Gemini 3.8 Flash has an introductory price of USD 0.75 per million input tokens and USD 3.75 per million output tokens until 31 December 2026; from 1 January 2027 it goes to USD 1.5 and USD 7.5 per million input and output tokens. This is the standard price for the global endpoint: the table on the page lists different amounts for non-global endpoints and for other service tiers.
• Voice, on Cloud Text-to-Speech. Chirp 3 HD voices cost USD 0.00003 per character (USD 30 per million characters) beyond the first million characters each month, which are not charged. The count includes spaces and line breaks. Gemini TTS models are charged by input text tokens and output audio tokens, and each second of audio is 25 tokens.
• Images, on Vertex AI. With Gemini 3.1 Flash Image, one generated image at 1K is billed as 1120 output tokens, about USD 0.067, and at 2K as 1680 tokens, about USD 0.101.
• Cloud editing, with Cloud Run. Jobs are billed for the entire lifetime of each instance that is started, with a minimum of 1 minute, and the price depends on the region. We give no figure here, because it depends on the CPU, the memory and the length of each video: see the Cloud Run pricing page (in Spanish).

Two details that change the sums. On Vertex AI you are only charged for requests that return a 200 response code: those that fail with 4xx or 5xx errors are not charged, but every result you discard and ask for again is. And if you pay in a currency other than the dollar, Google Cloud applies the prices listed for your currency, so any conversion you make yourself is only an estimate.

The full pages: generative AI pricing on Agent Platform (Vertex AI) (in Spanish) and Text-to-Speech pricing (in Spanish).

What does not show up on the invoice

The two items that are most often underestimated appear on no AI invoice:

• Your time. Choosing the idea, researching, reviewing the script, checking facts, approving scenes and preparing the title and thumbnail. If you value it at zero, any video looks cheap. Put an hourly price on it, even a rough one.
• Redos. A voice that does not convince, an image with an odd detail, an edit with a badly placed cut. If your way of working forces you to redo the whole video because of a change in one scene, the cost multiplies.
• Trials. Comparing several models or styles before you decide costs money. Record it separately, as a start-up investment for the channel, so that it does not distort the cost of normal videos.

How to measure the cost per video

Measuring well does not require complicated tools. A spreadsheet with one row per operation is enough:

1. Video and scene it belongs to.
2. Phase: script, voice, image or editing.
3. Provider and model.
4. Units: tokens, characters, images or minutes.
5. Estimated cost at the day's rate, in the provider's currency and converted to euros.
6. Actual cost, when the invoice arrives.
7. Whether it was a redo, and why.

When you close the video, add up by phase and add the hours you spent on it. After a few videos you will know which phase takes the money, which takes the time and what you redo most. That information is what lets you decide: perhaps a more expensive image model that gets it right sooner is worth it, or recording the voice yourself.

Cost versus results

The cost per video only makes sense alongside what the video delivers. Some ways to compare it:

• Production cost per thousand views: what the video cost to produce divided by its views and multiplied by a thousand. It lets you compare videos of different length and budget.
• Cost versus revenue, if the channel is monetised: what YouTube analytics shows for that video against what it cost.
• Cost versus business goals, if the channel belongs to a business: enquiries, contacts or website visits that come from the video.

Two cautions. A video's figures keep changing for weeks, so compare videos of a similar age. And the fact that a cheap video performs better than an expensive one does not prove the price is the cause: the topic, the title and the thumbnail weigh a lot. If you are looking to monetise, there is more context in how to monetise a YouTube channel made with AI.

Common mistakes

• Counting only the AI invoice and forgetting the hours.
• Looking at the channel's average cost and not at each video's: the average hides the videos whose cost shoots up.
• Not separating redos: without that data you do not know whether the problem is the model, the instruction or the script.
• Doing sums with an old rate: prices change, and some models already have their price announced for after the introductory period.
• Choosing the most expensive model for everything. For the script a fast model may be enough and it may pay to spend more on the voice or the images, or the other way round: compare it with your own videos before deciding.

If you are just starting, also read how to make YouTube videos with AI and how to choose the AI voice-over.

How the Estudio de Canales measures it

In the Estudio de Canales, the cost is not worked out by eye at the end. Every call to the AI has its cost recorded, linked to its provenance, and in addition:

• Each operation stores the original amount in the provider's currency, the currency, the reference exchange rate and the estimated euros. That dollar-to-euro rate is a reference assumption, not the day's market rate.
• What you see is the estimated cost. The actual cost is stated by the provider's invoice: the studio does not fill it in on its own and tells you how many operations still have no actual cost.
• If a model has no rate, no price is invented.
• You see it per video, per scene and per model, alongside the time each phase took and what had to be redone. The times are calendar times, from one milestone to the next, not hours of actual work.
• Regenerating a scene does not touch the others, and asking twice for exactly the same thing is not paid for twice.
• If you record your own voice in a scene, that narration has no provider and no AI cost.
• In results, it compares what the video cost to produce with what it is delivering, as cost per thousand views.

You can choose the model in each phase, and each one shows its reference price.

Sources consulted

Checked on 9 October 2026. We link to the original source so that you can consult the current text.

• Precios de la IA generativa en Agent Platform (Vertex AI). Google Cloud.
• Precios de Text-to-Speech. Google Cloud.
• Precios de Cloud Run. Google Cloud.