# AI video model specifications, with per-value provenance

**17 model families · 57 versions · 464 sourced values · 5 cross-field constraints**

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 a citation for the table — a citation for **each number**.

That is the whole point. Published comparisons of these models overwhelmingly restate vendor marketing, and marketing and API references disagree 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 on 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* `verified_on` in the corpus. Every value has been checked since that date. (Newest: `2026-08-31`.)
- **Licence: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)** — see [LICENSE.md](LICENSE.md).
- **Schema: [SCHEMA.md](SCHEMA.md)** — read this before using `values.csv`.
- Regenerated from the source database by `php artisan aidir:export-specs`. Nothing here is hand-edited.

## Files

| File | What it is |
|---|---|
| [`index.json`](index.json) | The canonical document: metadata, field dictionary, every model. |
| [`models/`](models/) | One JSON file per model family. Stable paths — link to these. |
| [`values.csv`](values.csv) | One row per model × version × field, provenance columns on every row. Opens in a spreadsheet. |
| [`constraints.csv`](constraints.csv) | Rules that couple two or more fields on one version. |
| [`SCHEMA.md`](SCHEMA.md) | Field dictionary and the shape of every object. |

## The three states — read this before you load it

A missing number means one of two completely different things, and this dataset refuses to conflate them.

| State | Count | Means |
|---|---:|---|
| `known` | 447 | Read off a vendor page. |
| `unpublished` | 17 | Somebody checked; the vendor publishes no figure. `source_url` points at the page that was read. |
| `absent` | 448 | Nobody has researched this field for this version. |

`unpublished` is a **result**, not a hole: "this vendor documents no frame rate anywhere" is a fact about the vendor and is worth knowing. `absent` is a hole, and is labelled as one. Do not map either to `false` or `0`.

## 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.

| Model | Versions | Known | Unpublished | Absent | Constraints | Verified since |
|---|---:|---:|---:|---:|---:|---|
| [Google Veo](models/google-veo.json) | 4 | 29 | 0 | 35 | 0 | `2026-08-20` |
| [Hailuo Video 海螺视频](models/hailuo-video.json) | 5 | 50 | 2 | 28 | 3 | `2026-08-20` |
| [Hunyuan Video 混元视频](models/hunyuan-video.json) | 2 | 13 | 2 | 17 | 0 | `2026-08-20` |
| [Jimeng AI 即梦 (Dreamina)](models/jimeng-dreamina.json) | 3 | 16 | 2 | 30 | 0 | `2026-08-20` |
| [Kling](models/kling.json) | 6 | 55 | 1 | 40 | 0 | `2026-08-21` |
| [LTX Studio](models/ltx-studio.json) | 3 | 24 | 0 | 24 | 0 | `2026-08-20` |
| [Luma AI](models/luma-ai.json) | 2 | 10 | 1 | 21 | 1 | `2026-08-20` |
| [Moonvalley (Marey)](models/moonvalley-marey.json) | 1 | 2 | 4 | 10 | 0 | `2026-08-31` |
| [Pika Labs](models/pika-labs.json) | 2 | 13 | 1 | 18 | 0 | `2026-08-20` |
| [PixVerse](models/pixverse.json) | 3 | 22 | 1 | 25 | 0 | `2026-08-20` |
| [Runway](models/runway.json) | 3 | 20 | 1 | 27 | 0 | `2026-08-20` |
| [Seedance](models/seedance.json) | 6 | 53 | 0 | 43 | 0 | `2026-08-20` |
| [Sora](models/sora.json) | 2 | 14 | 1 | 17 | 1 | `2026-08-20` |
| [Stable Video Diffusion](models/stable-video-diffusion.json) | 1 | 9 | 1 | 6 | 0 | `2026-08-20` |
| [Vidu 维度AI](models/vidu.json) | 6 | 47 | 0 | 49 | 0 | `2026-08-21` |
| [Wan 通义万相](models/wan-video.json) | 5 | 45 | 0 | 35 | 0 | `2026-08-20` |
| [Zhipu Qingying 智谱清影 (CogVideoX)](models/zhipu-qingying.json) | 3 | 25 | 0 | 23 | 0 | `2026-08-20` |

## Flagship versions at a glance

One row per family — its flagship version only. **This table is a summary and drops the provenance**; the files above do not. Use them for anything you intend to rely on.

Reading the marks:

- `—` — **absent.** Nobody researched this field for this version. It is not a zero and not a no.
- *not published* — **unpublished.** Somebody checked and the vendor documents no figure.
- `†` — the value carries a **`scope`**: a qualification that changes what it means in practice, such as a resolution that exists only at one clip length, or a price that applies to one tier. The summary has no room for the sentence; the JSON and the CSV carry it in full, and reading it is not optional.
- `⚠` — a **constraint** couples this cell to another on the same row. The two are not independently choosable, and `constraints.csv` says how.

| Model | Version | Status | Max clip | Max resolution | fps | Native audio | Price |
|---|---|---|---|---|---|---|---|
| Google Veo | Veo 3.1 | preview | — | 4K † | 24 | yes † | 0.4 USD per second † |
| Hailuo Video 海螺视频 | MiniMax H3 (Hailuo 3.0) | current | 15s | 2K † | 24 | yes † | 0.13 USD per second † |
| Hunyuan Video 混元视频 | HunyuanVideo 1.5 | current | 10s † | 720p † | 24 | *not published* | — |
| Jimeng AI 即梦 (Dreamina) | Dreamina app — Seedance 2.5 | current | 30s † | *not published* | — | yes | — |
| Kling | Kling 3.0 | current | 15s | 4K † | *not published* | yes † | 0.084 USD per second † |
| LTX Studio | LTX-2.5 | current | 20s † | 4K † | 50 † | yes † | 0.09 USD per second † |
| Luma AI | Ray3.2 | current | 20s † ⚠ | 1080p † | — | *not published* | 1.2 USD per clip † |
| Moonvalley (Marey) | Marey | current | *not published* | *not published* | *not published* | *not published* | — |
| Pika Labs | Pika 2.5 | current | 10s † | 1080p † | — | *not published* | 0.04 USD per second † |
| PixVerse | PixVerse C1 | current | 15s | 1080p † | *not published* | yes † | 0.01 USD per credit † |
| Runway | Gen-4.5 | current | 10s † | 720p † | — | yes † | 0.12 USD per second † |
| Seedance | Dreamina Seedance 2.5 | current | 30s † | 1080p † | 24 | yes † | 0.231 USD per second † |
| Sora | Sora 2 | legacy | 20s † ⚠ | 720p † | — | yes † | 0.1 USD per second † |
| Stable Video Diffusion | Stable Video Diffusion img2vid-XT 1.1 | current | 4s † | 1024x576 | 6 † | *not published* | — |
| Vidu 维度AI | Vidu Q3 Pro | current | 16s | 1080p † | 24 | yes † | 0.12 USD per second † |
| Wan 通义万相 | Wan 2.7 | current | 15s † | 1080p † | 30 | yes † | 0.15 USD per second † |
| Zhipu Qingying 智谱清影 (CogVideoX) | CogVideoX-3 (hosted) | current | 10s | 4K † | 60 † | yes † | — |

## How to cite

> 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.

If you are quoting a single number, cite **its own** `source_url` as well — it is in the data precisely so you do not have to take our word for it.

## What this dataset does not claim

- **It is not complete.** 448 field/version combinations are `absent`, and the coverage table shows where.
- **It is not a benchmark.** Nothing here measures output quality. These are documented capabilities and prices, read off vendor pages.
- **It goes stale.** Vendors ship weekly. `as_of` is the honest floor, and every value carries its own date so you can judge each one separately.
- **Prices are the vendor's list prices** on the date read, in the currency the vendor bills. Nothing is converted.