Coming from the OpenAI Videos API
Their Videos API was shut down on 24 September 2026 and they publish no one-to-one replacement. This endpoint takes the same requests on the same paths, so the base URL and the model id are the change.
sora-2 is refused rather than quietly served by a different model. Echoing a model name we did not run into your logs and onto your invoice would put a false record in the two places you would check it. The refusal names where to find the ids that work here.The change
Point the SDK at this gateway and name a model it serves. Everything else — the create call, the polling, the download — is the shape your code already builds.
from openai import OpenAI
client = OpenAI(
base_url="https://api.tokenify.dev/v1",
api_key="YOUR_TOKENIFY_KEY",
)
video = client.videos.create(
model="bytedance/seedance-2.0-mini", # not "sora-2"
prompt="a paper boat on still water",
size="1280x720",
seconds="8", # theirs took 4, 8 or 12
)
while video.status in ("queued", "in_progress"):
video = client.videos.retrieve(video.id)
content = client.videos.download_content(video.id, variant="video")
content.write_to_file("out.mp4")The models page lists what each one generates, or call GET /v1/models and read kind: "video".
What maps, and what does not
The models are not theirs, so some parameters have different ranges. Nothing is silently adjusted: a value this endpoint cannot honour is refused by name, with the acceptable ones in the message. A request answered as though your value had been used, when it was not, is the one failure worth more than a refusal.
| Field | Theirs | Here |
|---|---|---|
model | sora-2, sora-2-pro | A model this gateway serves. Their ids are refused with model_not_available |
prompt | Required | Required, unchanged |
size | 1280x720, 720x1280, 1024x1792, 1792x1024 | A pixel size that matches one of the model’s bands exactly. 1280×720 and 720×1280 are both 720p — the price follows the pixel count, not the orientation, and the orientation is carried through separately. Anything else is refused, including their 1024×1792: it is neither 720p nor 1080p, and substituting the nearer band would charge you for a frame size you did not choose |
seconds | 4, 8 or 12, defaulting to 4 | A whole number inside the model’s own range, which is in GET /v1/models. Omit it for the model’s minimum |
input_reference | An image used as the first frame | Not supported here yet, and refused rather than ignored: a disregarded first frame would produce a video of the prompt alone, charged in full, looking like the model had seen your image and dismissed it. Send a first frame through the model maker’s own dialect |
characters | Character ids | Not supported; refused by name |
status | queued, in_progress, completed, failed | The same four. A job cancelled here reports failed with a code, because their vocabulary has no cancelled and reporting completed would send you to download nothing |
progress | An integer | 0 while waiting, 100 when finished. The suppliers report nothing in between, so there is nothing honest to put there |
variant | video, thumbnail, spritesheet | Only video. The other two are derived assets this gateway does not generate, and asking for one is refused rather than answered with the video |
What it costs
Per output token, not per second of compute, and the count scales with the pixels, the frame rate and the length. There is no prompt charge. The rate depends on the resolution band, so the size you send decides the price — which is why an unrecognised one is refused rather than mapped to the nearest band.
The reply to a create carries estimated_output_tokens, which their API has no field for: a job that bills on settlement and quotes nothing at submit leaves you no way to know what you have committed to until it is spent. Their SDKs ignore the extra field.
The rates are per model and per band.
Errors
In their envelope — {"error": {"message", "type", "code", "param"}} — so a client that already handles errors from their chat or images APIs handles these without a branch. The error table lists every code this gateway sends and whether retrying helps.
A supplier’s own error text is never forwarded. It names them, and on these models it has been observed saying things that are not true of this gateway.