Sora API Shutdown — Migrating to Veo 3.1, a Code-Level Guide

OpenAI's Sora API is officially scheduled to shut down on 2026-09-24. From 2026-09-25 onward, requests to the Sora video endpoints are expected to return HTTP 410 Gone, and OpenAI has announced no official successor — every integration built on sora-2 or sora-2-pro needs a new home before that date. This guide walks the shortest path: migrating to Veo 3.1 (gemini-veo31) through byesu, an AI API gateway with OpenAI-compatible and Anthropic-native endpoints where one key covers Claude, GPT-5.6, Grok, Gemini and media models, billed pay-as-you-go with no subscription. Because byesu's video endpoint mirrors the async shape a Sora integration already uses, most migrations reduce to a base URL and model-name swap.

Shutdown timeline and what it affects

Date	What happens
Now → 2026-09-24	Sora API keeps working; this is your migration window
2026-09-24	Official shutdown date
2026-09-25 onward	Video endpoints expected to return 410 Gone; no official replacement inside OpenAI's API

Everything on the Sora video surface stops: task creation via POST /v1/videos, status polling, /content downloads, and any pipeline downstream of them. There is nothing to "upgrade to" within OpenAI's API — this is a full provider migration, not a version bump. Prompts also behave differently across video models, so budget time for a side-by-side test run rather than a deadline-day cutover.

Why Veo 3.1 is the natural replacement

• Visual quality: Google's flagship text-to-video model — high fidelity, cinematic camera work, lifelike motion.
• 1080p support: request 1920x1080 directly; Sora's public tiers topped out lower for most sizes.
• Image-to-video: first-frame reference via input_reference (up to 2 images) — the same field name Sora uses, so i2v code ports almost unchanged.
• Same async workflow: create task → poll status → download mp4, identical in shape to the Sora flow you already run.
• Official upstream via byesu: served over an official channel (not reverse-engineered or mirrored), with automatic refunds on failed tasks.

Request field mapping: Sora → byesu /v1/videos

Change the base URL from https://api.openai.com/v1 to https://byesu.com/v1. The auth header stays Authorization: Bearer sk-....

Sora request field	Sora values	Veo 3.1 via byesu	Migration note
model	sora-2 / sora-2-pro	gemini-veo31	One model replaces both tiers
prompt	free text	prompt — unchanged	Keep the scene + action + camera style
seconds	"4" / "8" / "12"	"4" / "6" / "8"	Map 12 → 8; 4 and 8 carry over
size	1280x720, 720x1280, 1792x1024, 1024x1792	1280x720, 720x1280, 1920x1080	Map pro landscape sizes to 1920x1080, portrait to 720x1280
input_reference	first-frame image	input_reference — same name	Multipart requests only; up to 2 images
Poll	GET /v1/videos/{id}	GET /v1/videos/{id}	Unchanged
Download	GET /v1/videos/{id}/content	GET /v1/videos/{id}/content	Unchanged, returns the mp4

Token must use the media group

Video models require a token in the media / media-gen group, otherwise you'll get a "no available channel" error. Create one in the console before testing.

Before / after: creating a task

# BEFORE — Sora (scheduled to shut down 2026-09-24)
curl https://api.openai.com/v1/videos \
  -H "Authorization: Bearer sk-OPENAI_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "sora-2", "prompt": "A hummingbird hovering over a red flower, macro lens, morning light", "seconds": "8", "size": "1280x720"}'

# AFTER — Veo 3.1 via byesu (two strings changed)
curl https://byesu.com/v1/videos \
  -H "Authorization: Bearer sk-YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model": "gemini-veo31", "prompt": "A hummingbird hovering over a red flower, macro lens, morning light", "seconds": "8", "size": "1280x720"}'

Full flow in Python (create → poll → download)

import requests, time

BASE = "https://byesu.com"          # was: https://api.openai.com
H = {"Authorization": "Bearer sk-YOUR_TOKEN"}

task = requests.post(f"{BASE}/v1/videos", headers=H, json={
    "model": "gemini-veo31",        # was: sora-2 / sora-2-pro
    "prompt": "Waves crashing on rocks, slow motion, cinematic, dusk light",
    "seconds": "8",                 # Veo 3.1 supports 4 / 6 / 8
    "size": "1920x1080",            # 1080p — an upgrade over most Sora sizes
}).json()

while task["status"] not in ("completed", "failed"):
    time.sleep(5)                   # tens of seconds to a few minutes is normal
    task = requests.get(f"{BASE}/v1/videos/{task['id']}", headers=H).json()

if task["status"] == "completed":
    mp4 = requests.get(f"{BASE}/v1/videos/{task['id']}/content", headers=H)
    open("out.mp4", "wb").write(mp4.content)

If a task ends as failed, the cost is automatically refunded — no manual claim needed.

Image-to-video (same input_reference field)

curl https://byesu.com/v1/videos \
  -H "Authorization: Bearer sk-YOUR_TOKEN" \
  -F model="gemini-veo31" \
  -F prompt="Slow push-in shot as wind stirs the leaves in the frame" \
  -F seconds="8" \
  -F size="1280x720" \
  -F input_reference=@first_frame.png

Polling and downloading are identical to text-to-video.

Billing comparison

Model	Per-second price	8-second clip
sora-2 (official list price)	$0.10/sec	$0.80
sora-2-pro (official list price)	$0.30–$0.50/sec by resolution	$2.40–$4.00
Veo 3.1 via byesu	$0.01875/sec	≈ $0.15

Veo 3.1 on byesu bills strictly per second: total = $0.01875 × seconds (6s ≈ $0.1125, 8s ≈ $0.15). The console shows your live billed price next to the crossed-out official list price, and failed tasks never bill. Top up with USDT, Alipay or WeChat Pay — pay only for what you generate, no subscription.

Migration checklist

1. Create a byesu token in the media / media-gen group.
2. Swap the base URL: https://api.openai.com/v1 → https://byesu.com/v1.
3. Replace the model name: sora-2 / sora-2-pro → gemini-veo31.
4. Remap seconds: 12 → 8 (4 and 8 unchanged).
5. Remap size: 1792x1024 / 1024x1792 → 1920x1080 / 720x1280.
6. Keep the create → poll → /content loop as-is; if you relied on webhooks, switch completion detection to polling GET /v1/videos/{id}.
7. Re-test your prompt library side by side — identical wording can render differently across models.
8. Test image-to-video with a multipart input_reference request.
9. Run both providers in parallel, then cut over well before 2026-09-24.

Related links

• Veo 3.1 model page (parameters & pricing detail)
• Media generation overview (image / video)
• Calling via the API (full request format)
• Quickstart (3-step setup)
• Online generation playground
• Console sign-up / create a token

FAQ

When does the Sora API shut down?

OpenAI's Sora API is officially scheduled to shut down on 2026-09-24. From 2026-09-25 onward, requests to the Sora video endpoints are expected to return HTTP 410 Gone, and OpenAI has not announced an official replacement video API — integrations built on sora-2 or sora-2-pro must migrate to another provider before that date.

What is the best alternative to the Sora API?

Veo 3.1 is one of today's leading mainstream text-to-video alternatives: strong visual quality, 1080p support, image-to-video via a first-frame reference, and an async task workflow. Through the byesu AI API gateway it is served as gemini-veo31 over an OpenAI-compatible /v1/videos endpoint, so a Sora integration migrates with a base URL and model-name swap.

How much code changes when migrating from Sora to Veo 3.1?

Very little. byesu's video endpoint mirrors the async shape a Sora integration already uses: POST /v1/videos creates the task, GET /v1/videos/{id} polls its status, and GET /v1/videos/{id}/content downloads the mp4. Point the base URL at https://byesu.com/v1, set model to gemini-veo31, and remap seconds and size to Veo's supported values.

How does Veo 3.1 pricing compare to Sora pricing?

On byesu, Veo 3.1 is billed per second at $0.01875/sec, so an 8-second clip costs about $0.15. For comparison, OpenAI's official list price at the time of writing is $0.10/sec for sora-2 and $0.30–$0.50/sec for sora-2-pro depending on resolution. Failed Veo tasks on byesu are automatically refunded.

Does Veo 3.1 support image-to-video like Sora's input_reference?

Yes, and it keeps the same field name. Send the create request as multipart/form-data and attach the first-frame image in the input_reference file field (Veo 3.1 accepts up to 2 images). The model animates the video starting from that image; polling and the mp4 download work exactly the same as text-to-video.