Find the six removed Sora identifiers in your repo, then move to Sume

To find what the Sora shutdown broke in your code, search the repository for six strings: the Videos API path and the five model identifiers that OpenAI listed for removal. OpenAI's deprecations page (read 2026-10-08) says the Videos API and the Sora 2 models were removed from the API on 2026-09-24, after a notice on 2026-03-24. The script below does that search. The second half of this post covers what you then change to call Sume's POST /v1/videos.

What OpenAI's page lists

The entry dated 2026-03-24 says OpenAI notified developers who use the Videos API and the Sora 2 model aliases and snapshots that they would be removed on September 24, 2026. The shutdown table has six rows. The recommended-replacement column shows a dash for each of them, so the page names no successor.

A scan that finds every hit

Dated snapshot names hide in config files, test fixtures, and environment templates, not only in application code. The scan reads every text file under the current directory and prints file, line, and the matched string. It lists the longest identifiers first, so a line with sora-2-pro-2025-10-06 reports that full name and not the shorter sora-2-pro. It also flags any existing /v1/videos path, which you must check by hand, because the host in front of that path decides which API it is.

Run it from the repository root with python3 scan.py, and treat the count on the last line as the size of the job. A count of zero is also useful: it means the removed identifiers are gone from the files you scanned, though not from environment variables or secrets managers that live outside the repo.

import asyncio, pathlib, re

IDS = ["sora-2-pro-2025-10-06", "sora-2-2025-12-08", "sora-2-2025-10-06",
       "sora-2-pro", "sora-2"]
PATTERN = re.compile("|".join(re.escape(i) for i in IDS) + r"|/v1/videos")
SKIP = {".git", "node_modules", ".venv", "dist"}

def scan(root: pathlib.Path) -> list[str]:
    hits = []
    for path in root.rglob("*"):
        if SKIP & set(path.parts) or not path.is_file():
            continue
        try:
            text = path.read_text(encoding="utf-8")
        except (UnicodeDecodeError, OSError):
            continue
        for n, line in enumerate(text.splitlines(), 1):
            for m in PATTERN.finditer(line):
                hits.append(f"{path}:{n}: {m.group(0)}")
    return hits

async def main() -> None:
    hits = await asyncio.to_thread(scan, pathlib.Path("."))
    print("\n".join(hits) or "no Sora identifiers found")
    print(f"{len(hits)} hit(s)")

asyncio.run(main())

What changes on the Sume side

Sume documents /v1/videos as agreeing field-for-field with the OpenRouter Video Generation API, not with OpenAI's Videos API. A client written from OpenRouter's docs works after you change the base URL and the key. A client written for OpenAI's Videos API needs its request and response handling checked against the Sume docs. Sume does not list Sora models, so a model id in your config must be replaced, not renamed.

Order of work

Run the scan first and keep its output as the migration list. Then read GET /v1/videos/models once and choose models from the supported_durations and supported_resolutions fields, since limits differ for each model. Last, add an Idempotency-Key to every submit that your code can retry. Do not resubmit a paid request only because a local timeout fired; poll the job id you already hold, as the jobs docs say.

Keep the first migration small. Pick one prompt from your old Sora library, submit it on one Sume model with an Idempotency-Key, poll to completed, download the file from the content URL, and read usage.cost from the poll response. That one round trip proves the key, the base URL, the model id, and the billing field before you change the rest of the code. Then repeat the scan to confirm that none of the six identifiers remain.