Every few weeks a new video model lands and the discourse resets to the same question: which one is best? It is the wrong question. On a real production brief the models are close enough that the deciding factor is almost never fidelity — it is what a failed generation costs you, and how many you expect to fail.
The economics of a retry
A generation you throw away costs the same as one you ship. That sounds obvious, and it changes everything about how a model should be chosen.
Suppose model A produces a usable clip 70% of the time at 40 credits, and model B does so 85% of the time at 90 credits. The naive read is that B is better. The real read is the cost per usable clip:
- Model A: 40 ÷ 0.70 ≈ 57 credits
- Model B: 90 ÷ 0.85 ≈ 106 credits
Model A is not worse. It is not even close to worse — it is roughly half the price for the same delivered output. What you paid for with model B was fewer attempts, and attempts are the cheap part.
Where the expensive model earns its keep
This inverts as soon as a retry stops being cheap. Three cases where it does:
- The input is expensive. If a generation consumes a shot you paid a crew to film, a failure costs more than credits.
- The iteration loop is human. A retry that needs an art director to look at it costs their time, not just the credits.
- You are near a deadline. Attempts take wall-clock time, and past a certain point you cannot buy more of it at any price.
Outside those, reach for the cheaper model and expect to run it twice.
What we actually do
We treat the first pass as reconnaissance. A cheap model, several variations, one prompt — the goal is not a finished clip but to find out whether the shot works at all. Composition problems, wrong energy, a subject that reads badly in motion: all of it shows up in a 40-credit draft as clearly as in a 400-credit one.
The expensive model is a finishing tool. Using it to explore is like shooting tests on film.
Once the shot is settled, the expensive model runs once, on a prompt that has already been proven. The hit rate on that final pass is far above 85%, because every uncertainty was resolved before it ran.
The version that generalises
The rule underneath all of this is not about video. It is that the cost of a generative pipeline is set by the number of attempts, not by the price per attempt — and the number of attempts is set by how much uncertainty remains when you press go. Spending money to reduce uncertainty pays. Spending it on fidelity you had not yet earned does not.