Model vendors announce capabilities at keynotes and ship a subset of them weeks later. The announcement gets indexed, quoted and repeated; the shipped spec sits in an interface nobody quotes. Resolution is the spec least safe to plan around, and reference limits are the one most often misread. The fix is checking the picker rather than the press release.
Every spec in this category comes from one of two places, and they don't agree as often as you'd expect.
How we found this out
We built a model comparison table from vendor documentation and got two of four rows backwards.
The table said Kling did native 4K and Veo capped at 1080p. Opening Hexcoded's own picker shows the reverse: Kling 3.0 Omni offers 720 and 1080; Veo 3.1 offers 720, 1080 and 4K. Both figures came from published vendor material, and both were wrong for the product a user actually reaches.
The same table said Seedance 2.5 accepted "up to 50 reference files." The real structure is up to 30 images, 30 actors, 10 videos and 10 audio, capped at 50 in total — and that ceiling drops to 8 on Wan 2.7. One number, presented as a platform capability, describing neither.
A third figure was worse because it sounded right. An internal reference described Seedance 2.5 as offering native 4K. That came from a vendor announcement; the picker offers 480p, 720p and 1080p. It was caught during a verification pass, which is the only reason you're reading this rather than a post repeating it.
None of the three came from carelessness. All three came from trusting an announcement over an interface.
An announcement is a statement of intent. An interface is a statement of fact. They get quoted interchangeably.
How does the gap open?
Five mechanisms, in rough order of how often they cause trouble.
| Mechanism | What happens | How to catch it |
|---|---|---|
| Cut before launch | A capability is announced, then doesn't make the release. The announcement stays indexed | Compare the keynote date with the interface. Trust the interface |
| Staged rollout | A tier ships weeks or months after launch, so both "yes" and "no" are true depending when you asked | Check whether a spec is described as available or planned |
| Plan-gated | The capability is real but not on your plan. 4K exists on Veo, and only above a certain tier | Check your own plan's ceiling, not the model's |
| Total versus per-type limits | A single reference number is quoted where the real structure is a total plus category caps | Look for whether a number is a total or one category's allowance |
| Renamed or folded | A previewed capability ships inside a different product, under a different name | Search for the capability rather than the version number |
Source: Hexcoded, September 2026, from verification passes on the current model roster. Mechanisms described are observed patterns rather than vendor policy.
Where this falls short. None of these involves anyone lying. A keynote is a legitimate statement of intent, and plans change between announcement and release for ordinary engineering reasons. The problem isn't the announcement — it's that the announcement outranks the interface.
Which specs are least reliable?
Not all specs fail equally. Three are worth extra suspicion.
Resolution. Announced most optimistically, and the one clients ask about. Seedance 2.5's 4K is the current case, and it isn't unusual. It's also the spec most likely to be plan-gated — 4K exists on Veo 3.1 and sits above the entry tiers, so "does this model do 4K" and "can I generate 4K" are different questions.
Reference and input limits. Often stated as a single number when the real constraint is a total plus per-category caps. Seedance 2.5 takes 50 assets in total, composed of up to 30 images, 30 actors, 10 videos and 10 audio — and the total itself varies by model, dropping to 8 on Wan 2.7.
Duration. Frequently quoted as a maximum without saying whether it's per generation or achievable through chaining. Veo 3.1 is locked to 8 seconds per pass in Create mode, and reaches well past a minute through Extend. Both numbers are true and they describe different things.
Where should you actually look?
In descending order of reliability.
The picker you're generating in. It describes what you can do today, on your plan, with your model. Nothing outranks it.
The platform's own model documentation. Useful for the things a picker doesn't surface — credit rates, prompt structure, mode availability.
An implementation contract. API reference or toolchain node documentation for an integration. Underrated, because code that gets parameters wrong doesn't run.
A vendor blog post or keynote. A statement of intent. Where it conflicts with any of the above, it loses.
A comparison post. Including ours.
- Open the model in the picker and read the settings before you promise anything
- Check your plan's resolution ceiling separately from the model's
- For any limit stated as a single number, find out whether it's a total or one category's allowance
- For duration, find out whether the figure is per generation or achievable through chaining
- Note the date you checked. This category changes monthly
What does this cost you in practice?
Nothing, until you've promised it.
A wrong spec in your head is harmless. A wrong spec in a client proposal is a schedule problem, and by the time you discover it the work is usually written around the assumption.
The three most expensive versions: promising a resolution your plan can't produce, building a reference-heavy workflow around an input limit that turns out to be a fifth of what you planned, and writing a scene for a duration that's only reachable through chained extension.
Do model documents help?
Yes, for a different set of things — and they carry their own caveats, which is worth seeing rather than assuming.
Hexcoded's own Seedance 2.5 reference publishes credit rates: 35 credits per second at 480p, 77 at 720p and 264 at 1080p, with cost scaling directly with duration and resolution tier. Seedance 2.0 runs 24, 53 and 129, plus a 4K tier at 264.
That document also carries an internal note flagging that the 1080p rate for Seedance 2.5 is derived from a single confirmed data point and recommending a second checkpoint before publishing extrapolated values.
Which is the right way to publish a number. The figure and its confidence level, together. It's also why the derived 4-second and 30-second 1080p costs aren't quoted here.
How do we handle this now?
Three commitments, stated so you can hold us to them.
Every model spec we publish is checked against the picker rather than against vendor documentation, and carries the date it was checked. If a row doesn't have both, it doesn't publish.
Where a vendor announcement and the interface disagree, the interface wins and the announcement gets no mention. A spec that was announced and cut isn't a caveat, it's absent.
And where a figure is extrapolated rather than measured, we say so or leave it out. The Seedance 1080p costs above are the working example.
Specifications and credit rates described here are current as of the publication date and change with model versions. This post is a secondary source about a primary interface — check the picker before relying on anything for a commercial delivery.
- An announcement is a statement of intent. An interface is a statement of fact. Quote the second
- We built a model table from vendor docs and got Kling and Veo exactly backwards on 4K. That's how easy it is
- Resolution is the spec announced most optimistically, and the one most likely to be plan-gated
- "Does this model do 4K" and "can I generate 4K" are different questions. Check your plan separately
- A single reference number is usually a total, not one category. Seedance 2.5's 50 is 30 images, 30 actors, 10 videos, 10 audio
- Reference ceilings vary by model. Fifty on Seedance 2.5, eight on Wan 2.7
- Duration figures may be per generation or reachable through chaining. Veo is 8 seconds a pass and well past a minute chained
- Check the picker on the day you commit a spec to writing, not the day you deliver
Because most articles quote a vendor's announcement rather than the interface. Announced capabilities get cut before launch, ship in stages, land behind a plan tier, or ship folded into a different product — and the announcement stays indexed and quotable while the shipped spec sits in a settings panel.
Not in the picker. It offers 480p, 720p and 1080p. Native 4K was announced and didn't ship. Any 4K described by a third-party platform is an upscale pass rather than native generation.
Up to 50 assets in total, composed of up to 30 images, 30 actors, 10 videos and 10 audio. So "50 reference images" is wrong twice over — it treats a total as a category and omits actors entirely. The total itself varies by model: Wan 2.7 allows 8.
Of the video models, Veo 3.1 — and it's locked to eight seconds per generation in Create mode. It's also plan-gated: 4K sits above the entry tiers, so the model supporting it and your plan permitting it are separate questions.
The picker you're generating in, first and above everything else — it describes what you can do today on your plan. Platform model documentation next, for credit rates and mode availability. Vendor blog posts and comparison articles last.
Look for a date, and treat the page's date and the capability's date as separate things. A page updated last week can still describe a capability announced last year that never shipped.
Specs with a source and a date
Creative Studio runs 30+ models on one credit balance, and every setting is visible before you generate — duration, resolution, aspect ratio and reference ceiling, per model. The interface is the spec sheet.
Open Creative StudioMore on model capability, access and rights in Models.