Usable rate is the benchmark nobody publishes
A model that wins every side-by-side can lose badly on credits per shipped shot. How to log first-pass usable rate and let it decide which model you route to.
A model at 30% usable and a model at 70% usable can look identical in a bake-off. Show a viewer the good outputs from each and they will split evenly, or prefer the prettier one. That comparison is measuring the ceiling. What decides your week is the floor — how often you get something shippable on the first try — and no public benchmark measures it, because it isn't a property of the model at all.
What every published board actually measures
Preference. Elo ratings on both of the big video boards come from the same method: humans shown two outputs and asked which they prefer. That's a real signal about a real thing — the aesthetic quality of output the model is capable of producing, aggregated across a general prompt distribution and a large pool of raters.
Two structural facts about that method matter here.
It's a comparison between survivors. Both clips in a pairwise vote exist and are being judged on their merits. Nothing in the method registers the case where a model produced nothing worth showing, because there is no "neither" button that feeds into the score in the way a discard feeds into your budget.
It averages over a prompt distribution that isn't yours. You don't shoot a general distribution. You shoot the same short list of shot types on repeat, and your usable rate on that list is what determines throughput. A board can hold prompts constant across models; it cannot hold your brief constant, because it's never seen it.
Both of those are fine. They're not flaws in the boards — they're the boundary of what a preference benchmark can answer. The mistake is treating the answer to "which produces prettier output" as the answer to "which should I route this shot to."
Be precise about which board you're citing, incidentally. Arena (arena.ai), which was LMArena until January 2026, and Artificial Analysis (artificialanalysis.ai) are unrelated organisations running separate vote pools on separate rating scales, so a number from one is not comparable to a number from the other. Their mid-2026 text-to-video orderings largely agree, but the exceptions are wide enough to matter — a model can top one board and be absent from the other's top ten, often because the two added it at different times. Versely's catalog ELO scores come from Artificial Analysis, matched by model name, which is what the quality-per-credit report plots against credit price.
Defining it tightly enough to be worth logging
A loose definition produces a number nobody trusts, including you in three weeks. Four decisions make it rigorous:
First-pass usable rate = generations you'd ship without regenerating, divided by generations dispatched for that shot type on that model.
What counts as an attempt. One dispatched generation that returned output. A job that errored and wasn't charged isn't an attempt. Neither is an editor preview — the editor's 480p preview pass is free and carries a short per-user cooldown, and it's a render of an assembly rather than a new generation, so it never enters the tally.
What counts as usable. Ships into the cut as-is, or with edit-step fixes only: trim, speed, colour, captions, audio mix. If making it work requires generating anything again, it is not first-pass usable.
What to exclude entirely. Prompt-authoring failures. If your first attempt came back wrong because the prompt was underspecified and you rewrote it, that's your draft, not the model's miss. Log it as a prompt revision and start the tally at the version you actually meant. Skip this exclusion and you'll spend a month measuring your own writing and blaming the catalog. Segment-level fixes are the other exclusion: a clip where one second is broken and you patch that one second counts as usable-with-fix, not a discard.
The log itself is six columns, one row per dispatched generation. A spreadsheet, not a system:
| Date | Model | Shot type | Attempt | Outcome | Credits |
|---|---|---|---|---|---|
| 12 Aug | Model A | dialogue CU | 1 | discard | 26 |
| 12 Aug | Model A | dialogue CU | 2 | usable | 26 |
| 12 Aug | Model B | dialogue CU | 1 | usable | 9 |
| 12 Aug | Model B | product pan | 1 | discard | 9 |
Outcome takes three values and only three: usable (first pass, no reroll), usable-with-fix (edit-step or segment patch only), discard. Anything more granular gets abandoned by Thursday.
Two weeks of ordinary work fills this in. You need roughly twenty attempts per model per shot type before the rate stops moving — below that you're reading noise, and the temptation to conclude something at attempt six is the main way this exercise fails.
The arithmetic that reverses rankings
Credits per usable output = the model's credit price ÷ its first-pass usable rate for that shot type.
Here's a sample log for one shot type — a dialogue close-up, historically one of the harder briefs for any model. The credit prices are real catalog prices. The usable rates are illustrative, because that's the entire argument of this post: they're a property of one team's brief and acceptance bar, not of the models, and copying my numbers would defeat the point of measuring yours.
| Model | Credits | Usable rate (sample) | Credits per usable |
|---|---|---|---|
| MiniMax H3 Text to Video | 26 | 35% | 74 |
| Flux 3 Text to Video | 9 | 70% | 13 |
| Kling O3 Pro Text to Video | 70 | 80% | 88 |
Read the middle column and the right column together. The cheapest model here isn't the value pick because it's cheap — it's the value pick because it's cheap and it lands. The most expensive one has the best hit rate and still finishes last, because an 80% rate cannot rescue a price that high. And the middle row wins by roughly 5.7x over the top row despite almost certainly losing a side-by-side on visual quality, which is precisely the outcome a preference benchmark cannot predict.
Now flip the shot type. On a static product pan, most current models are near-ceiling and the rates compress toward each other — at which point the ranking collapses back to sticker price, and the expensive model loses by even more. On a crowd scene with in-frame text, the rates spread further apart and the premium model can genuinely win. There is no single winner. There's a winner per shot type, which is why the shot type column in the log is not optional.
Why nobody publishes this
Because it isn't publishable. Usable rate is a joint property of four things: the model, the brief, the prompting style, and the acceptance bar of whoever is judging. Change any one and the number moves. A benchmark operator can standardise the first three and has no access at all to the fourth — your bar is set by your client, your platform, and how much time you have left.
That makes it structurally absent from every buying decision that relies on public information, which is most of them. It also makes it the highest-leverage number available to you specifically, because it's cheap to measure and nobody else can hand it to you.
Practical consequence: the leaderboard is a shortlist generator, not a decision. Use ELO scores and credit price or a value ranking to narrow thirty models to four. Then measure the four on your own work. When a high-ranked model shows a low usable rate on your brief, the usual explanation is that it's strong on aesthetics and weak on instruction-following — which is exactly the trade a preference vote rewards and a client brief punishes.
What to do with the number
Route by shot type, not by favourite. The output of two weeks of logging is a routing table: dialogue goes here, product goes there, crowds go to the one that can do crowds. That's a more valuable artefact than any model opinion.
Recompute on every release and every quarter. A point release can move a usable rate materially in either direction, and rates drift as your own prompting improves — which is a real effect and one worth separating from model change by noting your prompt version in the log.
Multiply it into budgets. Credits per usable, not credits per generation, is the figure a batch estimate should be built on. Ask the agent to total a planned shot list and check the balance before dispatching, then multiply by your measured rate — the reroll-rate view of the same data is the budgeting half of this argument, where usable rate is the model-selection half. Current catalog pricing sits on pricing if you want to sanity-check a shortlist before committing.
FAQ
Isn't usable rate just the inverse of reroll rate?
Arithmetically, close — a 50% usable rate is roughly a 2x reroll rate. They're used for different jobs. Reroll rate answers "what will this batch cost me." Usable rate answers "which model should this shot go to." Same log, two questions.
How many attempts before the number means something?
Around twenty per model per shot type. Below that the confidence interval is wider than the differences you're trying to detect, and concluding anything from five attempts is how teams end up with a permanent opinion built on one bad afternoon.
What if my usable rate is low across every model?
Then the problem is upstream of model choice — usually an underspecified prompt or a shot type nobody's model handles yet. Rates that are uniformly poor point at the brief; rates that vary widely between models point at the models. That divergence is itself the diagnostic.
Does a high usable rate justify a much higher credit price?
Only up to the ratio. A model at eight times the price needs better than eight times the usable rate to break even, and no model has an eight-fold headroom above a competent competitor because rates cap at 100%. That ceiling is why cheap-and-adequate beats expensive-and-excellent more often than intuition suggests.
Start the spreadsheet on your next batch. Six columns, three outcome values, two weeks — and you'll own the one number about these models that nobody can publish for you.