Gemini Omni Interactions API versus Sume /v1/videos: field mapping

Google's Gemini Omni call is POST /v1beta/interactions with a model and an input, and the video comes back inside the response as base64 data. Sume's equivalent is POST /v1/videos with model and prompt; it returns 202 with a job id and polling_url, and you download the MP4 from a content URL when the status is completed.

The two shapes

Google's minimal request is a model id of gemini-omni-1.1-flash and an input string. Video output appears in a steps array as a model_output with a video content item of MIME type video/mp4 and base64 data; the SDK exposes interaction.output_video.data. Sume follows the OpenRouter video shape: bare model id, prompt, optional duration, resolution, aspect_ratio, frame_images, input_references, and callback_url.

What changes in your code

The big change is the loop. With Google you read the video from the response; with Sume you poll GET /v1/videos/{id} and the statuses are pending, in_progress, completed, failed and cancelled. Cost is returned as usage.cost on completion. The single-reference rule also differs: a lone reference goes in input_references and conditions the whole clip.

A helper that behaves like output_video.data

This function returns raw MP4 bytes for a prompt, so the rest of an app that expects bytes from Google's SDK does not need to change.

import os, time, hashlib, json, requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}

def output_video_bytes(prompt, **extra):
    body = {"model": "gemini-omni-flash-1.1", "prompt": prompt, **extra}
    key = "map-" + hashlib.sha256(json.dumps(body, sort_keys=True).encode()).hexdigest()[:24]
    r = requests.post("https://api.sume.com/v1/videos", json=body, timeout=60,
                      headers={**H, "Idempotency-Key": key})
    r.raise_for_status()
    url = r.json()["polling_url"]
    while True:
        s = requests.get(url, headers=H, timeout=30).json()
        if s["status"] == "completed":
            return requests.get(s["unsigned_urls"][0], headers=H, timeout=120).content
        if s["status"] in ("failed", "cancelled"):
            raise RuntimeError(s.get("error", s["status"]))
        time.sleep(10)

data = output_video_bytes("A marble rolls down a chain-reaction track", duration=5)
open("out.mp4", "wb").write(data)
print(len(data), "bytes")

Migration steps

A base64 payload in a response is simple but heavy: a 10-second 4K clip can be many megabytes inside a JSON body. A content URL lets you stream the file to disk and keeps your JSON handlers small.

• Swap the model id to gemini-omni-flash-1.1.
• Replace the query-string key with a bearer header.
• Move to a poll loop, or set callback_url and verify the signature.
• Add an Idempotency-Key derived from the request body.

Features that do not map one to one

Google's guide lists extension up to 40 seconds and edit inputs of 10 seconds or less; on Sume, edit uses the Video Router's video_url field and no extend field is exposed. Google bills output by the token; Sume bills the provider list per second times 1.25. See the Video Router and Videos guide for the full contract.