⏳ Sora API Shutdown: Migrate Before September 24, 2026

The Sora API dies on September 24, 2026. That's the last date OpenAI will accept a request, and — per its own help center — the point after which Sora account data is permanently deleted. The consumer app already went dark on April 26. If you're one of the teams still routing production video through the API because the six-month runway made it easy to procrastinate, this is the guide we wished existed when we ported our own pipelines off Sora in July: a timeline recap, the honest reason it died, and a workload-by-workload migration matrix to the five models we actually run.

The shutdown timeline#

The announcement came on March 24, 2026, as an abrupt farewell post from the Sora team on X — no dates, just a goodbye and a promise of timelines to come:

The dates followed shortly after: April 26 for the app and web experience, September 24 for the API. That gave developers exactly six months — a civilized sunset by API-deprecation standards, but one that a lot of teams (including two we consult for) treated as infinite until the app actually vanished in April. The canonical reference is OpenAI's discontinuation FAQ, which also covers content export and the pro-rated refund window that opened in June.

One deadline inside the deadline: export is not migration. Your generated library disappears with your account data after September 24. Pull everything down first, then worry about where new renders come from.

Why Sora actually died#

The model was never the problem — our Sora 2 guide documents a genuinely strong generator, and its physics simulation still beats some of the models that outlived it. The business was the problem. A Wall Street Journal investigation, cited across the shutdown coverage, put Sora's operating burn at roughly $1 million per day while global active users fell from a peak of about 1 million to under 500,000. The reported Disney partnership — announced at $1 billion — was abandoned before money moved. Against an expected IPO, OpenAI chose coding tools and enterprise products over a money-losing consumer video app.

Why this matters for your migration: Sora didn't die because generative video is a dead end. Every alternative below is actively shipping updates. The lesson is platform risk — which is why our matrix favors models with multiple access routes over single-vendor lock-in.

By the numbers#

• September 24, 2026 — Sora API discontinuation; account data permanently deleted after this date (OpenAI Help Center)
• April 26, 2026 — Sora app and web shutdown (OpenAI Help Center)
• March 24, 2026 — shutdown announced via the Sora team's farewell post on X (TechCrunch; the announcement post itself)
• ~$1 million/day — reported compute burn (Wall Street Journal investigation, cited by Cheddar and The Decoder)
• ~1 million → under 500,000 — peak-to-trough global active users (same WSJ reporting)
• $1 billion — reported value of the Disney partnership abandoned before the shutdown (Variety)
• 6 months — total migration window developers got, announcement to API cutoff
• $0.15–$0.40/second — Veo 3.1 Fast vs. Standard via the Gemini API as of August 2026, per third-party pricing trackers; some list Vertex AI tiers higher, so verify against Google's live pricing page before committing volume

The migration matrix#

We picked five replacements and mapped them by workload, not by benchmark scores. All five are models we've run in production this year — pricing notes are hedged deliberately because several vendors reprice by the month.

If you'd rather watch someone stress-test the field before reading spec sheets, this is the most thorough hands-on comparison we found from the shutdown window:

Where we land overall: Veo 3.1 for parity, Seedance 2.5 for economics, MiniMax H3 for dialogue. Kling and Runway are specialist picks — superb at their lane, not general replacements. Our running scoreboard across all of them is the tested generator ranking.

Porting your Sora prompts#

The part nobody budgets time for. Sora trained everyone to write conversational scene descriptions — long, novelistic, camera implied rather than commanded. Here's how that style converts, based on re-running prompts from our Sora 2 prompt library against each model:

• → Veo 3.1: Nearly direct. Keep the natural-language description, add explicit audio cues ("distant traffic, no music") because Veo will invent a soundtrack if you don't specify one.
• → Seedance 2.5: Restructure into a shot list — subject, action, camera move, lighting, mood, sound — one clause each. Sora-style prose works but wastes the model's obedience; terse structure gets noticeably tighter results.
• → Kling 3.0: Keep the scene description but promote motion to the first sentence. Kling weights early tokens hard; a camera move buried in paragraph two often gets dropped.
• → Runway Gen-4.5: Compress. Runway rewards dense, keyword-forward prompts with explicit camera directives; Sora-length prose dilutes it.
• → MiniMax H3: Write dialogue and SFX inline where they happen in the shot, not as a footer. H3 generates synchronized audio from placement.

Budget roughly a day per pipeline for prompt conversion and two test batches. Ours took longer only where character consistency was involved — reference-image systems differ far more between models than text prompting does.

Migrate by September 1, not September 24#

Our strongest recommendation: treat September 1 as your real deadline. Three reasons from our own port. First, deprecating APIs degrade before they die — rate limits tighten and support evaporates, and we saw elevated Sora API error rates as early as July. Second, prompt conversion always surfaces a workload you forgot. Third, the export queue: everyone else will be downloading their libraries in the final week, and OpenAI has published no guarantees about export throughput under load. A three-week buffer costs you nothing; a missed cutoff costs your entire archive.

How we tested and picked#

We ported three production pipelines (text-to-video social clips, image-to-video cinematics, and a dialogue-driven character series) off the Sora API across July, running each candidate model against the same 20-prompt regression set we use for the Sora tool page. Picks reflect output quality, API ergonomics, and pricing stability as of August 2026 — not launch-week demos.

What we cut, and honest limits#

We left out open-weight local models — Wan-class checkpoints are real options if you own a beefy GPU, but they're a different guide, not a two-week migration path. We also cut the long tail of Sora-wrapper resellers still selling "Sora API access"; anything reselling the API dies with it on September 24. And a limit worth stating: pricing in this guide is hedged because it deserves to be. Veo's per-second rates have already shifted once this year, Seedance 2.5's credit pricing is still unpublished on some surfaces, and vendors reprice around exactly this kind of migration wave. Verify against live vendor pages before you commit volume — then move. The deadline is not hedged.