[rekram1-node]

Problem

OpenRouter lists 30 video-only models (output_modalities=["video"]) that the sync skips. /models reports "0" prices for them; the real prices are the pricing_skus in GET /api/v1/videos/models, which use about 45 SKU names. This follows #10593 (speech, transcription, embedding and rerank) and #10603 (images, merged). It's built on top of the image code and reuses its fetch helper, base-model resolution and price-entry builder.

Change

 fetchModels
   GET /models?output_modalities=text,decisions
-  GET /models?output_modalities=speech,transcription,embeddings,rerank
+  GET /models?output_modalities=speech,transcription,embeddings,rerank,video
+  GET /videos/models  → pricing_skus attached to each video model
 translateModel
   mediaType(model)
+    "video" → v2 doc: type video, base_model, [api.videos], cost from pricing_skus

Each file is v2_only = true, uses base_model, and authors [api.videos] (POST /api/v1/videos; @openrouter/ai-sdk-provider has videoModel).

SKU → CostV2

SKU pattern	CostV2
cents_per_second_output[_res], cents_per_video_output_second_res	video.output, unit = "second", cents ÷ 100
cents_per_second_video_continuation_res	same, when.input_video = true
duration_seconds[_with|_without_audio][_res]	video.output per second (USD), when.generate_audio
text_to_video_ / image_to_video_ / reference_ prefix	when.mode = text-to-video / image-to-video / reference-to-video (OpenRouter's generation modes)
video_tokens[_res][_with_video_input|_without_audio]	video.output per 1M tokens, when.input_video / generate_audio
cents_per_image_input, reference_images, reference_input_images	image.input, unit = "image"
_480p … _1080p, _2k, _4k	when.resolution with OpenRouter's request values (480p, 2K, 4K)

videoCost(skus)
  parse every SKU; any unknown or unrepresentable SKU → no cost
  base = plain SKU, else the first one
  for each other SKU, fewest conditions first
    if base + earlier overrides already give its price → skip
    else add override { when: its qualifiers, price }

# google/veo-3.1
[cost.video.output]
unit = "second"
price = 0.4

[[cost.video.output.overrides]]
when = { generate_audio = { equals = false } }
price = 0.2

[[cost.video.output.overrides]]
when = { resolution = { equals = "4K" }, generate_audio = { equals = true } }
price = 0.6

[[cost.video.output.overrides]]
when = { resolution = { equals = "4K" }, generate_audio = { equals = false } }
price = 0.4

generate_audio defaults to true for models with audio (OpenRouter docs), and billing follows that default. Live, Kling 3.0 Std billed $0.63 with no flag and $0.42 with generate_audio: false, which is 5 s × $0.126 and 5 s × $0.084. None of today's SKUs vary by duration.

Unrepresentable prices (no [cost] on these 2)

Model	SKU	Why
runway/aleph-2	minimum_cents_per_generation = 56	CostV2 has no minimum charge
black-forest-labs/flux-video-upscale	cents_per_megapixel_second_precise / _creative	megapixel × second isn't a unit

Also noted but not in the SKUs: Kling billed a 3.04 s clip as 5 seconds (see above). Its per-second prices hold from 5 s up.

Lab files

19 new, all v2-only:

models/
├── heygen/             heygen-video-1, avatar-iv
├── black-forest-labs/  flux-3-video, flux-video-edit, flux-video-upscale
├── minimax/            MiniMax-H3-Max
├── alibaba/            wan3.0-video, wan3.0-video-prime, wan2.7, wan2.6, happyhorse-1.0, happyhorse-1.1
├── bytedance-seed/     seedance-2.0-mini, seedance-1.5-pro
├── runway/             gen-4.5, aleph-2
└── kuaishou/           kling-v3.0-pro, kling-v3.0-std, kling-video-o1

Reused: minimax/MiniMax-H3 and MiniMax-Hailuo-2.3, bytedance-seed/seedance-2.0/-2.0-fast/-2.5, xai/grok-imagine-video/-1.5/-1.5-lite, and the three google/veo-3.1*-generate-preview labs (OpenRouter's google/veo-3.1*; the file keeps OpenRouter's name, "Veo 3.1"). OpenRouter's bytedance/ and kwaivgi/ prefixes map to bytedance-seed and kuaishou through a media-only table, so chat models under those prefixes don't change. Each new lab cites its release date and its input formats. New labs output video/mp4 only; generated sound is a track inside the MP4, so they add no generic audio output.

Out of scope

• generate-v2 still derives api from v1 fields, so the v2 output shows chat-completions instead of videos (known, deferred).

Verification

• bun packages/core/src/sync/index.ts openrouter --dry-run --no-issues: dev showed 0 created / 7 updated (live DeepSeek, MiMo and GLM price churn) / 566 unchanged. The 30 files come from a real sync run on this branch (30 created), with the 7 unrelated updates reverted; the dry run afterwards shows 0 created and only the same 7 updates.
• bun run v1:check: v1 output is byte-identical.
• bun validate: passes.
• bun test: same 11 failures as dev. 6 new OpenRouter video tests pass: type mapping, per second with resolution and continuation, audio and mode overrides, per-token SKUs, unrepresentable SKUs, and base_model/standalone files.
• Live: Grok Imagine Video 1.5 Lite 1 s 480p billed $0.02 (= base). Kling checks are above.
• v2 output (_experimental_v2.0_api.json): all 30 present with type = "video", video/mp4 and their costs.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.