AI video model specs, with a source on every value

Every value in this dataset carries the vendor URL it was read from, the title of that page, and the date it was read. Not one citation for the table — a citation for each number.

That is the difference worth downloading. Published comparisons of these models overwhelmingly restate vendor marketing, and vendor marketing disagrees with vendor API documentation far more often than you would expect: a model advertised at 4K that the API caps at 1080p, a family advertised at 3–10 seconds whose endpoints accept exactly 5 or 10, a resolution that exists only at one clip length. Where the two disagree here, the API reference is the value and the marketing claim is recorded in scope. Where two vendor pages disagree and neither is authoritative, the value is null and the conflict is the finding.

As of 2026-08-20 — the oldest verification date in the corpus, so every value in it has been checked since then. A dataset is exactly as fresh as its stalest claim.

Download

Per-model files live at /data/ai-video-model-specs/models/<slug>.json — stable paths, linked from the coverage table below.

Three states, kept apart

A missing number means one of two completely different things, and this dataset refuses to conflate them. Neither maps to false or 0.

Why a version, not a product

A live model family routinely runs several versions with incompatible specs at the same time. One takes any integer from 4 to 15 seconds while its sibling takes 6 or 10 and nothing between; one caps at 1080p while the previous generation offers 4K; one bills per clip and the next bills per second. Collapsing those into a single row is not a rounding error — it produces a number that is true of no shipped model. So the unit here is (model, version, field).

And a row can lie while every cell in it is true. One family offers 1080P at 6 seconds only, so a row reading “up to 10s” beside “up to 1080P” claims a clip that cannot be generated. Neither cell is wrong; the row is. Those couplings are recorded separately, in constraints.csv, naming the exact fields they bind — so a renderer can mark those cells rather than drop a footnote and hope.

Coverage

Read the last three columns together. A low known count beside a high absent count is a model nobody has finished researching, and the dataset says so rather than filling the gap with plausible numbers.

Licence and citation

Licensed under Creative Commons Attribution 4.0 International. Use it for anything, including commercially. Credit us and link the licence.

Individual facts are not copyrightable and we do not claim otherwise. That a given model accepts clips of 4 to 15 seconds is a fact about that vendor. What the licence covers is the compilation: the selection and arrangement, the schema, the written scope and note text, and the recorded findings about which vendor pages contradict each other.

RECATOOLS, AI video model specifications, with per-value provenance (as of 2026-08-20). https://recatools.com/datasets/ai-video-model-specs/ Licensed under CC BY 4.0.

Quoting a single number? Cite its own source_url too. It is in the data precisely so you do not have to take our word for it.

What this does not claim

• It is not complete. 448 field/version combinations are absent, and the table above shows where.
• It is not a benchmark. Nothing here measures output quality. These are documented capabilities and prices.
• It goes stale. Vendors ship weekly. Every value carries its own date so you can judge each one separately.
• Prices are vendor list prices in the currency the vendor bills, on the date read. Nothing is converted — where a vendor runs a separate mainland-China price list, that is a separate value, not the same price in another currency.