Sora 2 API Shutdown: Best Alternatives and How to Migrate (2026)
The Sora 2 API shut down on September 24, 2026. How to migrate a /v1/videos workflow to the TAVOA AI API: model mapping, request translation, and a checklist.

OpenAI removed the Sora 2 models and Videos API on September 24, 2026, six months after announcing the deprecation, with no named replacement. Products built on /v1/videos now need a new video backend and an updated request and response layer. This guide moves that workflow to the TAVOA AI API. One API account covers Seedance 2.5, Kling 3.0 and Wan 3.0.
What happened to the Sora 2 API?
OpenAI published a deprecation notice for Sora on March 24, 2026, and phased the product out in two stages. The Sora website and apps shut down first. The API followed on September 24, 2026. The Videos API and every sora-2 and sora-2-pro model, including dated versions, were removed from the platform, so any call to /v1/videos now returns an error.
Date
What changed
March 24, 2026
OpenAI publishes the Sora deprecation notice
April 26, 2026
The Sora website and apps close
September 24, 2026
The Sora 2 API and /v1/videos shut down
OpenAI's developer documentation still hosts the Sora guide, now marked as deprecated, and the deprecation notice names no replacement model. Teams with a video feature built on Sora now have two tasks: choose a new backend and rewrite the integration layer. This guide covers the integration side. For creators who used Sora through its interface, the tool-by-tool overview is in Sora alternatives in 2026. All details below were verified in September 2026.
Is there a drop-in replacement for the Sora API?
Sora requests do not work on other video APIs without changes. The closest replacement keeps the same architecture: an asynchronous job, a status check or webhook, and a file to download. With that architecture in place, the product logic around generation can stay, and the changes stay inside the integration layer.
Moving to another single-model API fixes the outage, but it recreates the same dependency: the whole feature again rests on one model from one vendor. TAVOA AI approaches this as an AI-native creative suite, where generation runs through one account across many models and tools. The TAVOA AI API is how TAVOA AI handles integration: one key, one prepaid balance in US dollars and one request lifecycle for more than 50 current video and image models.
Video models: the catalog includes Seedance 2.5, Kling 3.0 and Wan 3.0, as well as models from the MiniMax, LTX, PixVerse and Grok families.
Model switching: changing models means changing the endpoint path and the body fields. Polling, webhooks and file handling stay the same.
The full list of models is in the API catalog, and the product overview, from billing to organizations, is in Meet the TAVOA AI API.
Which TAVOA AI model fits a former sora-2 or sora-2-pro workload?
Seedance 2.5, Kling 3.0 and Wan 3.0 cover different Sora workloads, and all three run through the same TAVOA AI API integration.
Model
Fits
Max length per request
Max API resolution
Aspect ratios
Audio
Seedance 2.5
Longer clips with sound
30 s
1080p
16:9, 4:3, 1:1, 3:4, 9:16, 21:9
Yes
Kling 3.0
Multi-shot and higher-resolution workflows
15 s
Up to 4K, depending on endpoint
16:9, 9:16, 1:1
Wan 3.0
Long, reference-driven clips
16:9, 4:3, 1:1, 3:4, 9:16, adaptive
For reference, Sora 2 generated clips up to 20 seconds at 720p, and Sora 2 Pro reached 1080p, both in 16:9 and 9:16. Each model page in the API catalog lists its full parameter set, supported inputs and current rate.
Seedance 2.5 for long clips with sound. Audio is produced in the same pass, and separate endpoints handle image-to-video, references, edits and extensions.
Kling 3.0 for multi-shot scenes. It can produce several shots in one request. Standard, Pro, Turbo and 4K endpoints are available, and the 4K endpoints serve workloads that need output above the 1080p of sora-2-pro.
Wan 3.0 for reference-driven clips. It works through text-to-video, image-to-video and reference-to-video endpoints, with native audio.
How did the Sora API work, and what carries over to TAVOA AI?
The Sora API ran every generation in four stages. The TAVOA AI API follows the same order:
Submit a job. A Sora integration sent POST /v1/videos with the model, prompt, frame size and length, and got back a video object with an id and a status. On TAVOA AI, the request goes to the model's own endpoint and returns a request_id, a status_url and a cancel_url.
Wait in the queue. Sora jobs moved through queued and in_progress to completed or failed. TAVOA AI uses the same four statuses and adds two more final ones: nsfw for requests stopped by moderation, and canceled for requests withdrawn while still queued.
Poll or receive a webhook. Sora offered GET /v1/videos/{id} for status checks and the webhook events video.completed and video.failed. TAVOA AI offers the status_url for polling and a webhook attached to each request.
Download the file. Sora streamed the MP4 from /v1/videos/{id}/content, and download links stayed valid for up to one hour. TAVOA AI returns a video URL in the completed response, and the file stays available for at least seven days.
The overall flow stays the same: submit a job, wait for it to finish, save the result. Authentication, endpoint paths, request fields, response parsing, status handling, webhook setup and error handling all need updating.
How do Sora API settings translate to the TAVOA AI API?
The table below works as a translation guide: each row shows how one part of a Sora request is written for the TAVOA AI API.
What it controls
Sora API
TAVOA AI API
Authentication
Authorization: Bearer with an OpenAI key
Authorization: Key with a key ID and secret, server-side only
Where the request goes
POST /v1/videos, model named in the body
POST to the model endpoint, e.g. /bytedance/seedance-2.5/text-to-video
Request format
multipart/form-data or JSON
JSON
Clip length
seconds, sent as text ("8")
duration, sent as a number (8)
Frame size
size in pixels ("1280x720")
resolution ("720p") plus aspect_ratio ("16:9")
Audio
generated with the clip
a setting per model: generate_audio or sound
Start image
input_reference, matched to the frame size
image_url on image-to-video endpoints
Job identifier
video id
request_id
Completion notice
webhook set in project settings
hf_webhook parameter on each request
Finished file
download link valid up to one hour
video.url kept at least seven days
Step 1: Create an API account and key
The API is a separate product from a tavoa.ai plan, with its own account and billing. Setup takes four steps:
Sign up for an API account as an individual or as a company.
Add a payment method and top up the prepaid US dollar balance. The minimum top-up is $5.
Generate a key in the console at open.tavoa.ai. The key has an ID and a secret, and the full key is shown only once.
Store the key on your server. It authenticates every call and should stay out of browser and mobile code.
Official Python and TypeScript SDKs are available alongside plain REST. The full setup walkthrough is in How To Generate AI Videos Straight From the TAVOA AI API, and the reference documentation is at docs.tavoa.ai.
Step 2: Rewrite the generation request
Three things change in every request. The model name moves from the body into the URL. The Bearer header becomes a Key header. The form fields become a JSON body. A typical Sora request looked like this:
The same job on Seedance 2.5 through the TAVOA AI API:
Other Sora capabilities map to separate Seedance 2.5 endpoints. Media inputs go in as publicly accessible URLs.
Image references: input_reference moves to the image-to-video endpoint, which takes an image_url and an optional end_image_url for the last frame.
Character references: the reference-to-video endpoint accepts image, video and audio references in one request.
Extensions and edits: Sora's extensions and edits calls correspond to the video-extend and video-edit endpoints.
Step 3: Switch status checks and webhooks
Store the request_id as soon as a request is accepted. It identifies the job for polling, cancellation, webhook deduplication and support requests.
For polling:
Poll the status_url returned in the submit response, with backoff between checks.
Stop when the status reaches completed, failed, NSFW or canceled.
Cancel through the cancel_url while the request is still queued.
For webhooks:
Add your HTTPS endpoint, URL-encoded, as the hf_webhook query parameter when you submit the request.
Reply with any 2xx status within ten seconds, after recording the event.
Network failures and 5xx responses are retried for up to two hours. 4xx responses are treated as final.
Deliveries can repeat, so deduplicate by request_id and status. If delivery fails for good, read the result from the status endpoint.
Every webhook uses one envelope with four fields: request_id, status, error and payload. On success, the video URL sits at payload.video.url.
Step 4: Save the finished file and check the cost
A completed request returns the file at video.url. Output stays available for at least seven days, so the last step of the pipeline copies it to your own storage. Integrations that downloaded Sora files right away, because Sora links expired within an hour, keep that logic unchanged.
Cost planning works from published rates:
Per-second pricing: video models are priced per second of output, and the rate depends on settings such as resolution and audio.
Estimate before running: the estimate endpoint takes the same parameters as the generation request and returns the cost without starting a job.
No charge for failures: failed generations are not billed, and their cost returns to the balance automatically.
Hard stop at zero: the balance cannot go negative. New requests wait for a top-up, or auto top-up keeps production traffic running.
Current rates for every model and configuration are listed on the model pages in the API catalog.
How does TAVOA AI MCP fit into a Sora migration?
If Sora ran as a tool inside Claude or another AI agent, outside your own application backend, TAVOA AI MCP may fit better than the API. TAVOA AI MCP is how TAVOA AI plugs into agent workflows: Claude or any MCP-compatible agent connects through the server URL mcp.tavoa.ai/mcp and signs in with a tavoa.ai account, without an API key.
Beyond the comparison table below, four details matter:
ChatGPT: the TAVOA AI plugin connects ChatGPT the same way.
Veo 3.1: the website lineup available through MCP includes Google Veo 3.1, alongside Seedance 2.5 at up to 1080p, Kling 3.0 and Wan 3.0. Access to each model follows the account's plan.
Unlimited: generations through MCP spend plan credits at standard rates, since Unlimited access applies only on the website.
Assets: results appear in Assets on tavoa.ai, tagged with the MCP source, next to work made in the interface.
Criteria
TAVOA AI API
TAVOA AI MCP
Account
Separate API account
Your tavoa.ai account
Access
API key ID and secret
Sign-in, no key
Billing
Per generation from a prepaid dollar balance
Plan credits at standard rates
Model catalog
API catalog
Website lineup
Where results go
Output URL for your own storage
Assets on tavoa.ai
Built for
Generation inside your own product
Generation from an AI agent
Connecting Claude is covered in How To Generate AI Videos Straight From Claude with TAVOA AI's MCP. Other agents are covered in How to Generate AI Videos With AI Agents in 2026.
What should a Sora API migration checklist include?
Before switching production traffic, work through these steps:
Replace /v1/videos with the endpoint of the selected TAVOA AI model.
Replace Bearer authentication with a TAVOA AI API key ID and secret.
Map Sora fields such as seconds, size and input_reference to the new model's parameters, and check length limits: Kling 3.0 caps a request at 15 seconds.