The real economics of AI-assisted production

You have something that needs finishing. A product video with three seconds of dead air in the middle, a course module recorded in a room with a fridge in it, a set of landing-page images that have to look like they came from the same shoot. Every vendor selling generative tools quotes a price for that work, and the price sounds trivial. Ten cents a second reads like nothing next to a day rate.

The trivial-sounding price is per attempt, and attempts are the part nobody quotes. This guide is about where the money actually goes in AI-assisted production, which of your costs it can realistically remove, and what you have to keep doing yourself for the finished thing to be yours. It is not for productions running under a union agreement or a studio’s legal department, which have their own rulebook and their own lawyers. It is for the person who pays the bill and also does the edit.

The meter runs on attempts, not on what you keep

Generation is metered by output produced, not by output used. OpenAI lists sora-2 at $0.10 per second of 720p video, and sora-2-pro at $0.30 per second at 720p, $0.50 at 1024p and $0.70 at 1080p; the batch tier runs at half those rates [1]. Google’s Gemini API bills video generation on Gemini Omni Flash by the token, which it says “equates to an effective price of approximately $0.10 per second” for 720p output [2]. Images are cheaper and priced the same way: Gemini 3 Pro Image works out to $0.134 per 1K or 2K image and $0.24 at 4K, and Gemini 3.1 Flash Image runs from $0.045 for a 0.5K image to $0.151 at 4K [2].

Take a 30-second clip at $0.10 a second. That is $3.00. It is also $3.00 for the version where the hands are wrong, $3.00 for the one where the camera drifts, and $3.00 for the one you actually use. Your real unit cost is $3.00 multiplied by however many passes it takes, and that multiplier is a property of your shot, your reference material and your patience rather than anything on a pricing page. Budget the price of a keeper, not the price of a generation, and you will be roughly right instead of confidently wrong by a factor of ten.

The cheapest lever on that multiplier is deciding before you spend. Shot list, script, voice direction, the exact framing, the thing the viewer is supposed to notice: settle all of it in a text tool such as Claude, ChatGPT or Gemini, where a revision costs a fraction of a generated second, and only then start the meter [1][2]. Generators are bad at reading your mind and excellent at billing you for the attempt.

A subscription is the same meter with a cap on it

Consumer plans hide the meter behind credits, which makes them look like flat fees. They are not. Adobe’s Firefly Standard plan is $9.99 a month for 2,000 generative credits, Firefly Pro is $19.99 for 4,000, and Firefly Premium is listed at $139.91 as limited-time pricing against a regular $199.99 for 50,000 credits [3]. ElevenLabs runs the same structure for audio: the free tier gives 10,000 credits a month and no commercial licence, Starter is $6 a month for 30,000 credits and does include a commercial licence, and Pro is $99 a month for 600,000 credits [4].

Two things follow. The first is arithmetic: a credit is a currency whose exchange rate the vendor sets, so convert to dollars per finished image or per finished minute before you compare a plan against a freelancer’s quote or against a per-second API. A plan that looks generous at 2,000 credits is a different proposition once you know what one usable output costs in credits.

The second is contractual and catches people. The free tier of a voice tool that carries no commercial licence is not a cheap version of the paid tier, it is a different product with your intended use removed [4]. Check the licence attached to the specific tier you are on, not the one on the marketing page, before anything generated on it goes near a client deliverable or an ad.

The saving lands in post, not in the shoot

The most useful public data on where AI actually moves production costs comes from a company that files its numbers. In its Q2 2026 shareholder letter, Netflix wrote that “In 2026, GenAI workflows have been used in roughly 300 of our titles, with the largest concentration of work in post-production,” and that it is “increasingly leveraging these tools to deliver higher quality output more quickly and at a lower cost than traditional methods” [5]. That is a company describing its own operations to shareholders, which is a stricter setting than a product launch.

Note where the concentration sits. Not in writing, not in shooting, but in post: the stage where the expensive inputs, the people and the days and the locations, have already been paid for and the work left is finishing. The letter also names the second effect, which is easy to miss and worth more than the first: “In some cases, productions would have had to leave out key shots and sequences in the absence of GenAI technology,” citing enhanced crowds, historical battle sequences and worldbuilding establishing shots on specific titles [5]. Some of the value is a cheaper version of work you were going to do. The rest is work that was previously priced out of existence.

Your version of that is smaller and the same shape. Cleaning the fridge hum out of a recording, cutting filler words, extending a shot by a second so the cut lands, translating a module into a second language, replacing a background that was never going to be reshot. All of it operates on material you already captured, which means the model is finishing rather than inventing, and the attempt count stays low because there is a correct answer for it to converge on.

Feeding the model your own material lowers the attempt count

The one variable in the calculator you genuinely control is attempts per keeper, and the lever on it is reference material. Adobe’s Firefly Custom Models are trained on 10 to 30 JPEG or PNG images with a minimum width of 1,000 pixels, in one of three kinds: illustration style, photographic style, or a character [6]. The feature is in public beta for Creative Cloud customers on paid individual plans or Creative Cloud Pro and Pro Plus for teams, at no extra fee beyond the credits already in the plan, with training costing 500 credits per model and each generated image costing 20 [6].

Run that against the cheapest plan. Firefly Standard’s 2,000 monthly credits minus 500 for one training run leaves 1,500, which at 20 credits an image is 75 generations in the month [3][6]. Whether that is good value depends entirely on whether the tuned model gets you a usable image in two attempts where the untuned one took eight. Training is a fixed cost you pay once to reduce a variable cost you pay forever, which is the same trade as building a template, and it only pays back at volume.

There is a second reason to point these tools at your own material, and it survives every price change. Output built from footage, photographs and recordings you already own has an answer to the question of where it came from. That answer matters for the licence you can grant a client, for what you can register, and for the disclosure obligations further down this page. A model tuned on a folder of images whose provenance you cannot describe is cheap right up until somebody asks.

Ownership follows the human work, not the prompt

The U.S. Copyright Office addressed this directly in Part 2 of its AI report: “Given current generally available technology, prompts alone do not provide sufficient human control to make users of an AI system the authors of the output” [7]. The reasoning is that prompts “essentially function as instructions that convey unprotectible ideas,” and “at present they do not control how the AI system processes them in generating the output” [7]. Writing a very good prompt is not authorship, however long you spent on it.

What does count is stated just as plainly. “Copyright protects the original expression in a work created by a human author, even if the work also includes AI-generated material,” and where human-authored inputs are perceptible in the output, the human “will be the author of at least that portion of the output” [7]. Creative selection and arrangement qualifies too: a human “may select or arrange AI-generated material in a sufficiently creative way that ‘the resulting work as a whole constitutes an original work of authorship’” [7]. Modifications that meet the originality standard are protected as to “the material the human author contributed but would not extend to the underlying AI-generated content itself” [7].

In practice that puts a floor under the sensible workflow rather than the flashy one. Your footage and your recordings go in as inputs, so your expression is perceptible in what comes out. Your edit, your sequence, your choice of which of the eleven attempts survives, is the selection and arrangement. Keep the project file, keep the originals, and keep a short note of what you did, because the difference between an asset you can licence to a client and one you cannot is a record of human contribution rather than a feeling about it.

Disclosure is a line item now

Article 50 of the EU AI Act became applicable on 2 August 2026 [8]. Providers of systems generating synthetic audio, image, video or text must ensure outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated,” using solutions that are “effective, interoperable, robust and reliable as far as this is technically feasible” [8]. Deployers, which is what you are when you publish the output, must disclose that content has been artificially generated or manipulated, at the latest at the time of first exposure [8].

There is a carve-out worth knowing rather than relying on. Where the content forms part of “an evidently artistic, creative, satirical, fictional or analogous work or programme,” the obligation narrows to disclosing the existence of generated content “in an appropriate manner that does not hamper the display or enjoyment of the work” [8]. A marketing video for a real product is not an evidently fictional work. Treat labelling as a template decision made once, applied to every asset that reaches an audience in the EU, and priced into the job like colour grading, rather than an argument you have with yourself per file.

What still goes wrong

Every number above is a snapshot. Model names change, credit exchange rates change, promotional pricing expires, and a tier that includes a commercial licence today is a tier whose terms you should re-read next quarter. The prices here were checked on 4 September 2026 and are cited so you can check them again rather than trust them.

The attempts multiplier is the weakest part of any budget built this way, including one built with the calculator on this page. Nobody can tell you in advance how many passes a specific shot needs, and the variance between an easy background replacement and a shot with hands, text and a face in it is enormous. The calculator formalises your guess; it does not improve it. Run twenty seconds of the hardest shot in the job before you quote the whole job.

Unit price is also not total cost. Reviewing generated output takes human minutes that a per-second price does not include, storage and versioning grow faster than you expect, and rework after a client notices something is worse than never having generated it. And none of this is legal advice: the copyright analysis above is the Copyright Office’s own summary of its position [7], the Article 50 text is the regulation as published [8], and how either applies to your specific work is a question for someone qualified to answer it.