What one AI Short actually costs to make

Every tool in this category quotes you a subscription price. Almost none of them will tell you what a single video costs them to produce, which is a strange gap, because that number decides whether the subscription can survive contact with a customer who actually uses it.

Here is ours, with the important caveat that this is a dated production snapshot, not a timeless category benchmark.

The measured number

Across 20 entries in the production cost ledger, the mean was $0.18069 per Short and the median was $0.185364. The cheapest entry was $0.019548 and the most expensive was $0.4143.

That range is more useful than one rounded headline. A cached or straightforward subject can sit near the floor; a job with more sourcing, retries or discarded candidates moves toward the top. If you are planning capacity, the median describes the ordinary job and the maximum shows why a budget built around the cheapest render will fail.

Where the money goes

The current ledger records the all-in cost of each production. It does not provide an audited provider-by-provider allocation, so this article will not turn a few observed line items into a percentage and call it a fact.

What the production path does show is where variance enters: footage discovery, downloads, narration, model work and retries. Real-media sourcing is the least predictable part because one subject may yield usable archive material immediately while another needs more candidates before the quality and rights checks pass.

That inversion is worth sitting with if you are choosing tools. The parts being marketed to you as expensive and clever, the script and the voice, are commodities. The part nobody advertises, sourcing real footage with attribution and a usable licence, is where the actual work and the actual cost live.

Why the range is wide

Caching still matters: reusing a discovery result or a previously downloaded eligible clip removes work from the next run. So does rejecting weak candidates before downloading them. Those mechanisms explain why repeatable jobs can be cheaper, but the ledger snapshot above is the number we can reproduce today; it does not support a universal “cold” price or a fixed savings percentage.

What this means for the subscription

Some arithmetic that any tool in this category could publish and almost none do.

At a median production cost of about 19 cents, the render itself is only one part of what a subscription funds. The product also has to carry failed searches, retries, storage, the queue, support and the jobs that are refunded because they did not pass delivery checks.

It is also why there is no free tier. At these unit costs a free tier is affordable, which is exactly the trap. A free tier fills a product with people who will never pay while the queue that paying customers wait in gets longer.

What this means if you are building it yourself

You can. The pipeline is not magic and every component has a public API.

What you are actually taking on is the variance: finding footage that is real, downloadable, licensable and relevant, then proving it was not generated. That is the part that took the longest to build here and the part that keeps changing the cost after the code is built.

If you want a comparison point for your own build, the production timing snapshot contains 23 completed renders with a median of 697.5 seconds, about eleven and a half minutes from topic to finished file. Treat that as a dated observation, not a promise: source availability and queue state can move an individual job in either direction.