Sora 2 API Shutdown: Self-Hosted Video Alternatives (2026)

The Sora 2 API shutdown is real, and it's dated: September 24, 2026 is when the API stops working. OpenAI notified developers of the deprecation on March 24, 2026: the Videos API and every Sora 2 model alias, sora-2, sora-2-pro, and their dated snapshots, get removed from the API on that date, with no replacement model listed in the documentation (OpenAI). If you built a product on the Sora 2 API, this isn't a soft deprecation you can leave for next quarter. It's a hard cutoff, and OpenAI already gave you less runway than it looks like: the notice landed in March, which means half the migration window is already gone.

This post covers exactly what breaks, what OpenAI does with your data afterward, and the self-hosted stack we'd actually point a team at: Wan 2.2, HunyuanVideo 1.5, and where FLUX.2 and FLUX 3 fit into the picture, with real cost-per-clip numbers against Sora 2's published API pricing.

What the September 24, 2026 Sora 2 API Shutdown Actually Breaks

OpenAI's deprecation notice is specific about scope: the Videos API and all Sora 2 generation models, sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08, and sora-2-pro-2025-10-06, are removed from the API on September 24, 2026 (OpenAI). OpenAI's own language on the deprecations page: "On March 24th, 2026, we notified developers using the Videos API and Sora 2 video generation model aliases and snapshots of their deprecation and removal from the API on September 24, 2026." There's no successor model in the recommended-replacement column. The API isn't being upgraded to a new version, it's going away.

Timeline: App Shutdown (April 26) to API Shutdown (September 24)

Two separate shutdowns are easy to conflate, and they're five months apart:

The consumer app going dark in April didn't touch the API. It stayed fully operational, and still is as of this post, giving developers a five-month window from the app shutdown to the API shutdown to migrate (Pixo). If your team read the April headlines and assumed the API was fine indefinitely, it isn't. The clock has been running since March.

What Stops Working, and What Happens to Your Data After

After September 24, 2026, every Sora 2 API call returns a 410 Gone error, and OpenAI deletes all Sora account data permanently, with no recovery window (Pixo). Anything you haven't already pulled off OpenAI's servers and stored yourself is gone the day the cutoff hits.

There's a second trap worth flagging: features that worked in the Sora app don't automatically carry over to the API, and never fully did. A moderator on OpenAI's developer community addressed this directly when a developer asked about reusable avatar handles: "Something that works in the Sora app, like reusable avatars/handles, does not automatically mean the same object or handle can be used through the API" (OpenAI Developer Community). If any part of your product quietly depends on app-only behavior rather than documented API endpoints, audit that now, not during the September scramble.

AI Video Generation on GPU Cloud: Self-Hosted Alternatives to Sora 2

Losing an API you don't control is the argument for owning the pipeline instead. Running AI video generation on GPU cloud means renting bare-metal GPU capacity, deploying an open-weight model yourself, and paying for compute time rather than per-second API calls that a vendor can shut off on five months' notice. The tradeoff is real: you take on model ops, and no open-weight model today matches Sora 2's native audio track. But you own the model, the weights, and the uptime.

Three model families cover most of what a Sora 2 migration actually needs.

Wan 2.2 for Text-to-Video and Image-to-Video (the Last Open-Weight Wan)

Wan 2.2 is the model to deploy if you want the closest thing to a drop-in self-hosted replacement for general text-to-video and image-to-video work. It's also, as of this year, the last Wan release with public weights: Alibaba's Tongyi Lab shipped Wan 2.5, 2.6, and 2.7 as API-only products, with no GitHub repo and no HuggingFace checkpoint for any of them. We covered that pattern in detail in deploying Wan 2.7 on GPU cloud, and the short version is: if a product page says "Wan 2.7," it's a hosted API call, not something you can run on your own hardware. Wan 2.2 is what you actually deploy.

Wan 2.2 is a 27B-parameter Mixture-of-Experts diffusion transformer (14B active per step). It needs roughly 40-48GB VRAM at 480p with FP8 quantization, and 65-80GB at 720p, which makes an 80GB H100 PCIe the practical minimum for 720p output with a tight margin, and an H200 the comfortable choice for 10-second 720p clips. The full ComfyUI and diffusers setup, including Docker configuration and weight download commands, is in our original Wan 2.1/2.2 GPU deployment walkthrough.

HunyuanVideo 1.5 for Lightweight Self-Hosted Video

If your team doesn't have H100-class hardware in the budget, HunyuanVideo 1.5 is the model built for that constraint. Tencent released it as an 8.3B-parameter model with weights and inference code on November 20, 2025, followed by official LoRA training scripts on December 5, 2025 (LocalAIMaster). It runs on as little as 14GB VRAM with model offloading enabled, and comfortably on 24GB consumer cards once you include the text encoders and the built-in super-resolution upscaler. It generates natively at 480p and 720p, with a distilled upscaling network that gets output to 1080p as an integrated step rather than a separate pass.

That's a real jump from the earlier, heavier HunyuanVideo line, which needed 60-80GB VRAM and 15-25 minutes per 5-second 720p clip on an H100 (Spheron: GPU cloud for video AI 2026). HunyuanVideo 1.5's footprint fits comfortably on an L40S (48GB) or an A100 80GB, both well below H100 pricing.

One catch worth knowing before you commit to it commercially: HunyuanVideo 1.5 ships under the Tencent Hunyuan Community License, which explicitly excludes commercial use in the EU, UK, and South Korea, and requires a separate license from Tencent above 100 million monthly active users (LocalAIMaster). If your product serves those regions commercially, get legal to read the license text before you build on it. For a hands-on setup comparison against Wan 2.2 image-to-video, see our image-to-video GPU deployment guide covering LTX, Wan, and Hunyuan.

Where FLUX.2 and FLUX 3 Fit, and Where They Don't Yet

FLUX.2 from Black Forest Labs is worth knowing about here, but it's not a video model. It's a 32B image-generation transformer, downloadable and self-hostable right now: roughly 64GB VRAM at BF16, about 32GB at FP8 (fitting an H100 PCIe or A100 80GB), and as low as 13GB for the distilled FLUX.2-klein-4B variant, which we cover in the FLUX.2 production deployment guide. Its role in a Sora 2 migration is keyframes, thumbnails, and reference images that feed into Wan 2.2's image-to-video mode, not a video generation replacement on its own.

FLUX 3, which Black Forest Labs launched July 23, 2026, is closer to a genuine Sora 2 competitor on paper: a single model trained jointly on image, video, audio, and robot-action data, with FLUX 3 Video generating clips up to 20 seconds in one pass, including native synchronized audio (BFL). The problem is access. Only FLUX 3 Video and FLUX 3 Action are live, and both require gated early access to selected partners. FLUX 3 Image is still rolling out, and FLUX 3 Dev, the eventual open-weight release, has no announced date, license, or spec sheet. We go through the full breakdown in FLUX 3 GPU requirements. Watch it, but don't plan your September migration around it.

Cost and Quality Tradeoffs: Sora 2 API Pricing vs Owning the Pipeline

Sora 2's per-second pricing is straightforward on paper and gets expensive fast at scale. The question that actually matters for a migration decision is what the same clip costs to generate on your own rented GPU instead, and whether that math holds up at your volume.

Sora 2 / Sora 2 Pro Per-Second Pricing vs Self-Hosted GPU Cost Per Clip

OpenAI's published API rates, per second of generated video (OpenAI):

At those rates, a 20-second sora-2 720p clip runs $2.00; the same 20 seconds on sora-2-pro at 1080p runs $14.00. Batch processing halves both.

Here's what comparable clips cost self-hosting Wan 2.2 on Spheron, using current live GPU rates:

Pricing fluctuates based on GPU availability. The prices above are based on 08 Aug 2026 and may have changed. Check current GPU pricing → for live rates.

Spot pricing on self-hosted Wan 2.2 beats Sora 2's standard API rate at both clip lengths in this table, and on-demand runs close to even. None of that accounts for idle GPU time, which is the real variable: a rented instance bills whether or not it's generating, so at low, bursty volume the API's zero idle cost can still win on paper even where its per-second rate looks worse. If you'd rather stay on a hosted API but leave OpenAI specifically, Fal.ai is worth a look as a middle ground, per-second billing without the infrastructure ops, just not from a vendor with a shutdown clock already running.

Migrating Off Sora 2 Before September 24: A Practical Checklist

Five months sounds like a lot of runway until you map out what actually has to happen before the cutoff. Work through this in order:

1. Download everything you need now. Once September 24 hits, OpenAI deletes all Sora account data with no recovery window. Pull every generated video, prompt, and metadata record you might need before that date, not the week of it.
2. Audit for app-only dependencies. Anything your product relies on that only worked in the now-defunct Sora app, reusable avatar handles being the documented example, was never guaranteed to work through the API. Build strictly against documented API endpoints going forward.
3. Pick your model by volume and resolution, not by hype. Wan 2.2 covers general text-to-video and image-to-video closest to what Sora 2 did. HunyuanVideo 1.5 covers the same ground on cheaper hardware if your resolution needs top out at 720p-to-1080p-upscaled. Neither generates audio natively, so plan a separate TTS or audio pass if your product needs sound.
4. Provision GPU capacity. Rent H100 SXM5 or B200 capacity for Wan 2.2 at 720p, or step down to an L40S or A100 for HunyuanVideo 1.5. Spheron's docs cover instance provisioning if you're setting up bare-metal GPU access for the first time.
5. Run the cost math at your actual volume, not a single-clip estimate. The API-vs-self-hosted tradeoff isn't universal, it's the same shape of decision we walked through for LLM inference in GPT-6 vs self-hosted LLMs: low, bursty volume tends to favor staying on an API; consistent daily volume tends to favor owning the GPU.
6. Ship the migration before the cutoff, not on it. Give yourself buffer for the model quality gap. Wan 2.2 and HunyuanVideo 1.5 are good, but they're not Sora 2 on every axis, and finding that out in production on September 23 is the failure mode this checklist exists to prevent.

The Sora 2 API goes dark on September 24, 2026 with no replacement model announced, so the migration window is now, not later. Wan 2.2 and HunyuanVideo 1.5 run today on Spheron's bare-metal GPU instances with no per-second billing and full root access.

Rent H100 SXM5 → | View all GPU pricing →