API

Every model on Remerged, from your own code. A job made with the API is the same job the site makes: the same models, the same library, the same email when it is ready.

Start

1. Make a key on your account page. It is shown once.
2. Add credits there. Any plan can, Free included, and on the API a Free account uses every model. One credit is one US cent; the API never runs a tab.
3. Send the key as a bearer token. Check it with /account, price a job with /quote, make it with /jobs, then poll.

curl https://remerged.ai/api/v1/jobs \
  -H "Authorization: Bearer rm_live_…" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A hummingbird hovering at a red flower, slow motion",
    "model": "minimax-h3",
    "durationS": 5,
    "aspect": "16:9"
  }'

# 201 {"id": "…", "ids": ["…"], "credits": 35}

curl https://remerged.ai/api/v1/jobs/<id> -H "Authorization: Bearer rm_live_…"
# {"job": {"status": "generating" | "upscaling" | "ready" | "failed", "video_url": …}}

Price

Paid plans get 10% off the API. Prices below are before that discount; /models and /quote give your own. Paid from your balance and fixed when the job is made. Models on our own GPU are priced to include starting a machine for you; partner-hosted renders that come in under their quote are credited back when the video is delivered. Five seconds at each model's own lowest quality (the mode a job gets when it names none):

Upscaling to 1080p or 4K is added per second. Ask /quote for any other combination.

Endpoints

All under https://remerged.ai/api/v1. JSON in and out. Errors are usually {"error": "code", "message": "…"}, but some carry only a sentence in error: go by the HTTP status first.

The same, machine-readable: /openapi.json (OpenAPI 3.1, the job body generated from the schema we validate with) and /llms.txt for language models.

The job body

• prompt, model (an id from /models), durationS, aspect (16:9, 9:16, 1:1, 4:3, 3:4, 21:9 or adaptive), audio.
• mode: the render quality, optionally with an upscale, e.g. 720p, 720p-4k, 480p-1080p. /models lists each model's. Leave it out for the model's own lowest quality: 480p on Seedance, 720p on the others. Upscaling is never added unless you ask for it.
• image_urls: public HTTPS images to start from, as URLs or {"url": "…", "role": "first_frame"}. Or upload with /uploads and pass images: [{"key": "uploads/…", "role": "reference"}].
• seed, variations (up to four; not both, since one seed makes one video), webhook_url.
• Per model (/models says which): negativePrompt, what to keep out (Wan 2.6, up to 500 characters); promptExpansion: false, Wan 2.6 renders your prompt exactly as written; draft: true, a Seedance 2.5 480p draft you can finish at 1080p as the same take with /jobs/{id}/finish. Wan 2.6 also renders 1080p natively.
• MiniMax H3, LTX-2.5 and Wan 2.6 take a prompt and a starting image; MiniMax H3 also a last frame, and Wan 2.6 a soundtrack of your own as reference_audio (3 to 30 seconds). The Seedance models take up to 30 pictures on 2.5 and 9 on 2.0, reference video and audio, first and last frames, and characters (characters need Pro). A Seedance video starts from frames or from references, not both: a first or last frame cannot go with reference pictures, video, audio or characters. Reference clips and audio tracks run at least 2 seconds; audio up to 15 seconds on Seedance 2.0 and 30 on Seedance 2.5, and on Seedance 2.0 an audio track needs a picture, a person or a clip beside it. Give each clip's length as seconds: then anything a model cannot take is refused with unsupported_inputs before you are charged. Without it we measure the file, and a clip of the wrong length fails the job, refunded in full.
• Prompts go to the model exactly as written, except on Wan 2.6, which expands each prompt itself and may cut a video made from words into several shots. Refer to pictures as Image 1, Image 2… in the order you send them (images first, then image_urls), and to clips and audio as Video 1 and Audio 1. Pictures for Seedance are 300 to 6000 pixels a side and at most 2.5 times as wide as tall (or tall as wide); for Wan 2.6, 240 to 7680. Wan 2.6 renders no 21:9 from words. A seed is 0 to 2147483647. The help centre has the prompting guides.

The API makes videos. Pictures, extending a clip, music and Studio are on the site only for now.

When it is ready

Poll /jobs/{id} every ten seconds or so, or pass a webhook_url: when the job is ready or has failed we POST its body there, signed with your key's webhook secret (shown with the key when you make it) as X-Remerged-Signature: sha256=<hex HMAC of the body>. You also get the usual email. Video URLs last an hour; ask again for a fresh one.

Limits

• 60 requests a minute per key.
• Jobs in flight at once: Free 2, Standard 4, Pro 8, Max 16.
• Five keys per account. A key is made and revoked on the account page only.
• No credits, no job: a 402 carries a link to add some.