# OpenRouter video catalog — 2026-09-10

Evidence for `182.C.10.01` (capability vocabulary and its provenance). The
question this measurement answers is not "what models exist" but "which of the
things a creator asks for does the catalog actually tell us", because every
entry the catalog is silent about is an entry a registry can only get from a
probe or a dated human record — and reading silence as `false` is the failure
mode `libs/contracts/src/human-video/capability-vocabulary.ts` exists to stop.

## Method

Key sourced at run time from `~/.config/oshun/env-master.env` through a
throwaway `/tmp` script, never written into the repo. Verified by length only.

```
GET https://openrouter.ai/api/v1/videos/models
Authorization: Bearer $OPENROUTER_API_KEY
```

- `200`, 26578 bytes, **28 models**, read 2026-09-10.
- Raw response: `videos-models.json` (pretty-printed; field order and values
  otherwise unmodified).

A first attempt used `GET /api/v1/models` — 436 models, output modalities
`['audio', 'image', 'text']`, **no video** and none of the video capability
fields. The video catalog is a separate endpoint; the general model list cannot
answer any question in this section.

## Every record carries the same 17 keys

```
id  canonical_slug  hugging_face_id  name  created  description
supported_resolutions  supported_aspect_ratios  supported_sizes
supported_durations  supported_frame_images  upscale_factor  creativity
generate_audio  seed  pricing_skus  allowed_passthrough_parameters
```

Five are identity/description (`id`, `hugging_face_id`, `name`, `created`,
`description`); `hugging_face_id` was `null` on all 28. Twelve bound behaviour.
`unmappedCatalogFields()` in the contract is checked against exactly this key
list, so a field OpenRouter adds later surfaces as a failing test rather than a
silently ignored capability.

## A present field is often a null field

Counted across the 28 records:

| Field                     | null | Reading                                                    |
| ------------------------- | ---- | ---------------------------------------------------------- |
| `upscale_factor`          | 27   | genuinely absent; only `flux-video-upscale` upscales       |
| `creativity`              | 27   | same model, same reason                                    |
| `supported_sizes`         | 9    | resolution tier offered with no enumerated `WxH` list      |
| `supported_frame_images`  | 4    | no frame-image slots advertised                            |
| `generate_audio`          | 4    | catalog declines to state; not a statement of "no audio"   |
| `supported_durations`     | 3    | e.g. `heygen/avatar-iv`, whose duration follows the script |
| `seed`                    | 3    | catalog declines to state                                  |
| `supported_resolutions`   | 2    | `flux-video-upscale`, `runway/aleph-2`                     |
| `supported_aspect_ratios` | 1    | `flux-video-upscale`                                       |

So the vocabulary gives each catalog-sourced entry a `fallbackWhenAbsent` of
`probe`, except `upscale_factor` and `creativity` where null IS the answer.

Booleans, where stated: `generate_audio` true 18 / false 6 / null 4; `seed` true
16 / false 9 / null 3. Frame images: 24 models advertise `first_frame`, 15 of
those also `last_frame`, and 9 are first-frame-only (`alibaba/wan-2.6`, `-3.0`,
`-3.0-prime`, `happyhorse-1.0`, `-1.1`, `runway/gen-4.5`, both
`x-ai/grok-imagine-video*`, `minimax/hailuo-2.3`). Only `first_frame` and
`last_frame` were seen as slot names.

Resolution values are vendor tier labels, not dimensions: `480p`, `720p`,
`768p`, `1080p`, `2K`, `4K`. `768p` exists because `minimax/hailuo-3-max` says
so, which is why the tier list is read from the catalog rather than assumed.

## What the catalog does not carry, for any model

No field anywhere in the response states:

- how many references a model accepts (**no** reference-count limit exists in
  any record) — probe;
- whether a model accepts supplied speech audio — the only trace is an `audio`
  name in `allowed_passthrough_parameters` on `alibaba/wan-2.7` and `-2.6`,
  which is a parameter name and not a semantic;
- whether a model can edit or extend an existing take — the traces are a
  `keyframes` passthrough on `runway/aleph-2`, vendor prose about
  instruction-guided editing on `minimax/hailuo-3`, and the priced SKUs
  `cents_per_second_video_continuation_720p` / `_1080p` on
  `black-forest-labs/flux-3-video` plus `video_tokens_with_video_input` on the
  Seedance models. A priced continuation is real corroboration and still says
  nothing about how it is invoked;
- retention, data handling, or geographic restriction — dated manual records.

`pricing_skus` is heterogeneous, as the phase snapshot said: per second
(`minimax/hailuo-2.3`), per second by tier (`alibaba/wan-3.0`), with/without
audio (`google/veo-3.1`), by generation mode (`alibaba/wan-2.6`,
`kwaivgi/kling-v3.0-*`), per `video_tokens` (all Seedance), per megapixel-second
with a precise/creative split (`flux-video-upscale`), and `runway/aleph-2` adds
a `minimum_cents_per_generation` floor of 56 cents. Cost estimation needs the
normalizer in `182.C.10.05`.

## Re-verification of the six lane rows in the phase snapshot

Checked field by field against this response; **all six unchanged** from the
2026-09-09 table.

| Lane candidate               | Frames       | Durations | Resolutions       | `generate_audio` | `seed` | Passthrough                                                                                                                   |
| ---------------------------- | ------------ | --------- | ----------------- | ---------------- | ------ | ----------------------------------------------------------------------------------------------------------------------------- |
| `alibaba/wan-2.7`            | first + last | 2–10      | 720p, 1080p       | true             | true   | `negative_prompt`, `prompt_extend`, `audio`, `ratio`, `last_image`, `video`, `videos`, `images`                               |
| `alibaba/wan-2.6`            | first only   | 5, 10     | 720p, 1080p       | true             | true   | `negative_prompt`, `enable_prompt_expansion`, `shot_type`, `audio`, `size`                                                    |
| `bytedance/seedance-1-5-pro` | first + last | 4–12      | 480p, 720p, 1080p | true             | true   | `watermark`, `req_key`                                                                                                        |
| `bytedance/seedance-2.5`     | first + last | 4–30      | 480p, 720p        | true             | true   | `watermark`, `req_key`, `output_format`                                                                                       |
| `google/veo-3.1`             | first + last | 4, 6, 8   | 720p, 1080p, 4K   | true             | true   | `personGeneration`, `aspectRatio`, `negativePrompt`, `conditioningScale`, `enhancePrompt`                                     |
| `heygen/avatar-iv`           | none         | null      | 720p, 1080p       | false            | false  | `voice_id`, `voice_settings`, `motion_prompt`, `expressiveness`, `fit`, `remove_background`, `background`, `caption`, `title` |

Two details worth carrying forward: `bytedance/seedance-2.5` advertises
`generate_audio: true`, so the "no audio" in its snapshot row means no audio
**input**, and the flag must still be set deliberately for an exact-audio
request. `google/veo-3.1` exposes `aspectRatio` as a passthrough alongside the
top-level `aspect_ratio`, so the two paths must not both be set.

The catalog has also moved since 2026-09-09 in ways that do not touch the lane
rows: 28 entries now, including `minimax/hailuo-3-max` (768p, 5–15 s),
`bytedance/seedance-2.0` (up to 4K, 5040x2160 among its 26 sizes),
`bytedance/seedance-2.0-mini` and `-fast`, `runway/aleph-2`,
`black-forest-labs/flux-3-video` and `flux-video-upscale`,
`x-ai/grok-imagine-video-1.5`, `alibaba/happyhorse-1.0`/`-1.1`, and
`kwaivgi/kling-video-o1`. That churn is the argument for `182.C.10.03`: this
table is a dated observation, not a constant.
