Nine specs a model page should publish

Most model launch posts publish a demo reel, a comparison table against the previous version, and a price. None of those answer the question a working team actually has: can this model do the specific job already on the calendar. That decomposes into about nine facts, and the gap between how many a model page publishes and how many it should is a fair proxy for how much testing you are about to do.

Here is the list, what each one decides, and what it means when it is missing.

Group one: the capability gates

1. Duration ceiling, and the allowed set. Not the maximum. The set. Video models do not accept a continuous range, they accept published values, and the sets differ enough that "max 15s" and "3s to 15s in one-second steps" are very different products. Veo 3.1 allows 4s, 6s and 8s. Hailuo 2.3 Pro allows 6s and 10s. Kling 3 Turbo runs 3s to 15s. LTX 2.3 Text to Video Fast goes to 20s but in two-second steps starting at 6.

The floor matters as much as the ceiling. A model whose shortest allowed duration is 5s cannot give you a 3s beat, and trimming afterwards does not recover the credits you spent on the seconds you threw away.

Omission means: you cannot plan a shot list. In Versely's own catalog, 24 of 146 video models publish no duration list, so it gets resolved at dispatch rather than at planning time.

2. Native frame rate. The rate the model actually generates at, and which alternatives it exposes. This one bites a week later rather than immediately, because a single clip at the wrong rate looks fine and a timeline of mixed rates does not.

Versely renders at 25 fps by default. A model publishing 24, 30 and 60 is offering you a choice; a model publishing nothing is offering you whatever it does natively, and you find out by measuring the output.

Omission means: conversion artefacts you will blame on the model's quality. Two thirds of the video models in the catalog publish no frame rate list at all, which is the single worst coverage figure on this list.

3. Resolution path. Three separate facts hiding under one word. What resolution does it generate natively? What tiers can you select? And is the top tier native or an upscale pass bolted on?

The third question is the one nobody answers and the one that changes your workflow. A native 4K model and a 1080p model with an upscaling stage produce different artefacts, and only the first resolves detail that was never there. Pixel count grows with the square of the label, so the tiers are not evenly spaced in time or credits either.

Omission means: you plan around a number that turns out to be a marketing ceiling. 36% of video models in the catalog publish no output resolution ceiling, and 30% publish no quality tier list.

Group two: control and reproducibility

4. Audio, and which kind. Four states, not two. No audio. Native audio generated in the same pass. Audio as a separate pass you can turn off. Audio-driven generation, where a supplied track drives the picture. Different costs, different render times, different roles in a pipeline.

If the clip is going under a voiceover and a music bed, generated audio is discarded and you waited for it anyway. If the model's audio is the deliverable, the sync quality is the whole review.

Omission means: you cannot tell whether to budget for a separate audio stage. Of 146 video models in the catalog, 67 declare audio, 18 declare no audio, and 61 declare nothing either way. That last group is 42%, and it is the group where you find out by generating.

5. Reference slots. How many references, of what type, and how they are addressed. "Supports references" is not a spec. "Up to 9 images, 3 videos and 3 audio references" is. So is "one image, used as the first frame".

This decides whether a model can hold a character or a product across a sequence, which is the difference between a model you can build a series on and one you can make single clips with. Reference images and reference-to-video are different input modes and a page should say which it means.

Omission means: you cannot plan character consistency, which is usually discovered on shot four. 87% of video models in the catalog carry no reference configuration, so this is an industry-wide gap rather than anyone's individual failure.

6. Seed control. Whether a seed is exposed, whether passing the same one reproduces the same output, and whether that holds across resolution and duration changes.

Reproducibility is the whole basis of controlled comparison. Without a stable seed you cannot isolate a prompt change from sampling noise, which means every A/B you run on that model is really an A/B/C where C is luck.

Omission means: no controlled testing. Worth being honest here: this is not a field in Versely's model records either. The generation tools accept a seed parameter, but the per-model page does not publish whether that seed is honoured deterministically, and it should.

Group three: terms and economics

7. Usage terms. Commercial use permitted or not. Territory restrictions. Attribution requirements. What the provider says about likenesses and about training on your inputs.

None of this is a quality judgement and all of it is a gate. A model that cannot legally carry your paid placement in the market you are running in is not a candidate, whatever its leaderboard position, and the answer lives in the provider's published terms rather than the model card. Same discipline as usage rights on creator content: silence in a licence gets read narrowly, not generously.

Omission means: legal review lands after production instead of before it, which is the expensive order.

8. Retake behaviour. What happens when you run it again. Three sub-questions: does a reroll at a fixed seed return the same clip, how much does output vary at different seeds, and does the model expose any explicit variation control.

Seed variance is the spec that most often explains why a model that demoed brilliantly disappoints in production. A model producing one great clip in five is not a cheaper version of a model producing five good ones, it is a model with a reroll budget attached, and that budget belongs in the comparison.

Omission means: your effective cost per usable clip is unknown until you have run thirty of them.

9. Credit cost, and what drives it. The number, and the shape. Most video models are billed on what they actually render, so length, output resolution and whether an audio pass runs all move the figure. A single price is therefore a floor, a ceiling, or a lie, and a page should say which. Versely's catalog publishes a range per model along with the driver, and the agent's cost estimator returns per-item and total credits against your balance before anything dispatches.

Worth stating plainly because it changes how you read every other spec on this list: there is no free tier here and no free allowance. Every generation costs credits, including the ones you run to fill in the gaps a model page left. The one exception is the editor's 480p preview pass, which is free and carries a short per-user cooldown; the final export is charged once regardless of how many clips went into it.

Omission means: you cannot forecast a batch, and batch forecasting is where most budget overruns start.

What the omissions collectively mean

Read across those nine and a pattern emerges. The specs that get published flatter: resolution ceilings, maximum durations, leaderboard positions. The specs that get omitted constrain: allowed value sets, seed determinism, reference limits, retake variance, licence terms.

That is not usually deception. It is that the flattering specs are single numbers and the constraining ones are sets, conditions and caveats. But the practical consequence is the same either way: the fields a model page leaves blank are the fields you will spend credits discovering. Budget accordingly. A model with a thin record is not necessarily a bad model, it is a model with a testing cost attached, and that cost is real even when the credits per clip look attractive.

The corollary is a buying heuristic. Between two models of similar apparent quality, prefer the one whose record is complete: you reach production faster and your comparisons are reproducible. Ranking data has the same problem, with 149 of the catalog's 296 model records carrying none at all, so a leaderboard position exists for roughly half the catalog. The quality-per-credit report plots that half.

FAQ

Which of the nine matters most?

Duration set and reference slots, because those two are hard gates rather than quality gradients. A model that cannot hit your length, or cannot hold your character, is disqualified regardless of how good it looks. Everything else on the list is a trade you can make consciously.

Can I get these facts if the page does not publish them?

Partly. Duration, aspect ratio, resolution and reference limits are usually recoverable from the model's input schema at dispatch, which is what the agent's schema lookup returns, one entry per provider route. Seed determinism, retake variance and usage terms are not recoverable that way. The first two you measure, the third you read in the provider's terms.

Is a sparse record a sign of a new model?

Sometimes, and that is the charitable reading: records fill in over the weeks after a launch as fields get confirmed. It is still a cost to you meanwhile, and "new" is not a reason to plan a client deliverable around a model whose duration set nobody has written down.

How should I compare two models when both records are incomplete?

Hold everything constant and generate. Same prompt, same duration, same aspect ratio, same resolution tier, three seeds each, and estimate the credits before you dispatch so the comparison has a budget. That is more work than reading a comparison page, which is exactly the cost the missing specs imposed on you.