Sora API Shuts Down September 24, 2026: What to Do

TL;DR — what happens on September 24, 2026

Short on time? Seven points. Each one is unpacked below with dates and snapshot names.

• On September 24, 2026 OpenAI removes the Videos API and every Sora 2 alias and snapshot from the API. Requests to them stop returning video.
• Developers were notified on March 24, 2026 — exactly six months of notice. This is a planned deprecation, not an outage.
• Five model names go dark: sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06.
• OpenAI named no recommended replacement — the replacement column on the deprecations page is empty. There is no successor to Sora inside OpenAI.
• The Sora web and mobile apps were switched off earlier, on April 26, 2026. You can still export your clips at sora.chatgpt.com/sunset via the Export button — do it before September 24.
• The real replacements this autumn: Seedance 2.5, Kling 3.0, Gemini Omni Flash, MiniMax H3, Veo 3.1 and open-weights Wan 2.7.
• In Twin AI six of them share one credit balance: a short clip is 600 credits on Hailuo 2.3, Kling and Seedance and 900 on Veo 3.1.

Exactly what is being removed: endpoints and snapshots

The confusion comes from “Sora” meaning three different things: the sora.com app, generation inside ChatGPT, and the Videos API for developers. September 24 only touches the third one, because the first two are already gone.

OpenAI’s deprecation list puts the September 24, 2026 date next to six entries. Check your code against it — if a single line matches, that is the line that breaks:

• The Videos API itself — the create and retrieve endpoints, not just a model name.
• sora-2 — the alias most integrations were built on.
• sora-2-pro — the higher-quality alias.
• sora-2-2025-10-06 — the first public snapshot.
• sora-2-2025-12-08 — the December snapshot.
• sora-2-pro-2025-10-06 — the pro snapshot of the same date.

The full Sora shutdown timeline

Sora was not switched off in one move — it took four steps over six months. That is why so many teams only learn about the deadline now: the first three steps never touched them.

• March 24, 2026 — OpenAI announces that Sora is being discontinued and, the same day, sends developers the formal deprecation notice for the Videos API and the Sora 2 snapshots.
• April 26, 2026 — sora.com and the Sora iOS and Android apps stop accepting new generations. For ordinary users, Sora ends here.
• September 24, 2026 — the Videos API and all Sora 2 snapshots are removed. Anything calling them from code breaks.
• After September 24 — only your archive remains: the files you exported in advance. There is no way back to your generations on OpenAI’s side.

Your checklist before September 24

Three jobs. The first one is urgent because it has a hard deadline; the other two have slack.

First, export your videos. Everything you generated in Sora sits at sora.chatgpt.com/sunset behind an Export button. Download the archive and put it in your own storage. There will be no way to recover it afterwards: the app is already dead and the API is going.

Second, find every place your code calls Sora. Not just the obvious generation call — there is usually a hardcoded model name in a config, a retry path, a webhook that receives the finished video, and a database table where duration and resolution are stored in Sora’s formats. The fastest sweep is a grep for “sora-2” and “videos”; the snapshot names are predictable.

Third, pick a replacement and run your real prompts through it. Not ten pretty demos — twenty typical production jobs. The gap between models shows up on your boring repeated scenes, where face stability, small-detail legibility and predictable clip length matter, not on “a cat in a hat”.

What to replace Sora 2 with: the autumn 2026 field

In the six months Sora spent dying, the video market reshuffled completely. Artificial Analysis Video Arena blind rankings for August 2026 put Gemini Omni Flash in front at 1245 Elo, with MiniMax H3 right behind at 1242 and Seedance 2.0 at 1225. Veo 3.1 sits only eleventh at 1098, below all four Kling 3.0 variants — yet it remains the only top model with an official 4K tier and always-on synchronised audio. Judge by capability, not by rank:

• Seedance 2.5 (ByteDance) — the most general-purpose move off Sora. Its API opened on July 16, one request carries up to 30 seconds and up to 50 references (2.0 allowed 12). In the image-to-video-with-audio arena, 2.0 holds first place at 1196 Elo.
• Kling 3.0 (Kuaishou) — native 4K, multi-shot storyboarding and the cleanest camera control on the market. All four variants outrank Veo 3.1 in the overall arena.
• Gemini Omni Flash (Google) — the overall arena leader. 720p with native audio at roughly $0.10 per second, but clips cap at 10 seconds at 24 fps. Chosen for iteration speed and conversational frame editing.
• MiniMax H3 — second in the arena and the cheapest way to push high content volume. The direct successor to the Hailuo line that many creators moved to back in spring.
• Veo 3.1 (Google DeepMind) — if your Sora pipeline involved audio, dialogue or lip sync, this is the only direct equivalent. Plus an official 4K tier.
• Wan 2.7 (Alibaba) — open weights, the self-host leader. Alongside it, LTX-2.3 ships 4K plus audio under Apache 2.0, and HunyuanVideo 1.5 renders in 75 seconds on a single RTX 4090. Insurance against depending on someone else’s shutdown decision twice.

One shot in three lighting treatments — this is how you actually pick a model for your look

Which model for which job

A short map so you do not have to guess. Product and commercial work where material texture matters — Seedance 2.5 or Kling 3.0. Animating a still photo while keeping the composition — Kling 3.0, its camera control is the tidiest. Dialogue, voiceover, lip sync — Veo 3.1; nothing else matches it on audio. Twenty short social clips a day — MiniMax H3 or Hailuo 2.3, where per-clip cost decides, not ten Elo points. A fully offline pipeline on your own hardware — Wan 2.7.

When the call is close, testing is cheaper than arguing. Twin AI has a /compare mode for exactly this: the same prompt goes to several models at once, the results land side by side, and you pick the winner by eye. It is what Sora power users used to do by hand across five tabs and five subscriptions.

What it costs in credits

The Sora API billed per second and required a foreign card. Twin AI needs no subscription — credits are pay-as-you-go, and the same balance covers images, video and chat. Here is the cost of one generation at default settings — a 5-6 second clip, with Veo setting its own length — so you can size the migration budget straight away:

• Hailuo 2.3 — 600 credits for a 6-second clip. The fastest queue for covering a stream of short clips.
• Kling 2.6 — 600 credits for 5 seconds. The previous Kling generation, still a solid workhorse.
• Veo 3.1 Fast — 600 credits per generation. Audio and lip sync at mid-market cost.
• Wan 2.7 — 600 credits for 5 seconds. What you would self-host, without the GPU.
• Kling 3.0 in standard mode — 600 credits for 5 seconds (Pro +240). Camera control and multi-shot.
• Seedance 2.0 — 600 credits for 5 seconds. The all-round frame-quality replacement for Sora.
• Veo 3.1 — 900 credits per generation. Top quality with sound.
• A longer clip, audio switched on or pro mode are surcharges on top of the base price — the composer always shows the total before you launch.

For developers: rewriting off the Videos API

The migration is not hard, but it is not a one-line swap either — the Videos API had its own async mechanics and they do not match other vendors.

• Do not hardcode the new model. The mistake that brought you here was a single hard dependency on one vendor. Put the model name in config behind a routing layer.
• Rework the waiting logic. Sora handed back a job and a status; Seedance, Kling and Veo each have their own polling and webhook formats, and longer timeouts — a 30-second Seedance 2.5 clip takes far longer than a 5-second Sora frame.
• Recompute your length and resolution limits. Gemini Omni Flash caps at 10 seconds at 720p, Seedance 2.5 holds 30 seconds, Veo 3.1 has an official 4K tier. Any validator carrying Sora’s limits is now lying to you.
• Revisit your prompts. Sora’s grammar (subject + camera + style + duration) ports to Seedance and Kling almost unchanged. Veo wants an explicit audio cue for lip sync. MiniMax prefers a short dense prompt — long Sora-style paragraphs confuse it.
• Build in a fallback model. When one queue stalls, the generation goes to the next. That is how routing works in Twin AI: one key, one balance, several models behind it — see /models.

If you never wrote any code

If you simply made clips in the Sora app, migrating is one action: open /create/video, pick a model, paste the same prompt. There is nothing to relearn — the phrasings that worked in Sora work here.

One thing Sora never had: the Twin Elo leaderboard at /models/leaderboard. It is built from blind A/B comparisons by real users rather than vendor claims, so you can sort models by quality before spending a single credit.

And free credits on signup are enough to try several videos before any payment, with no card required.

Common mistakes when leaving Sora

What usually goes wrong for teams that left the migration until September:

• Not exporting the videos. The Export button at sora.chatgpt.com/sunset works until the shutdown — after that the files exist nowhere.
• Waiting for an “official OpenAI successor”. There is none: the replacement column on the deprecations page is empty, and that is a decision to exit consumer video, not an oversight.
• Swapping one hard dependency for another. If you just typed Seedance where Sora used to be, you repeated the same mistake and postponed it.
• Picking by overall leaderboard rank. Arena leader Gemini Omni Flash caps at 10 seconds; if you need a 30-second clip, first place buys you nothing.
• Testing on pretty demo prompts. The gap between models appears on your routine scenes, not on showcase ones.
• Leaving a sora-2 call in production “until the last minute”. On September 24 it stops answering without further warning — the warning already came, in March.