OpenAI's Sora API Shuts Down Sept. 24: Where Its 4 Biggest Jobs Go Next

OpenAI's Videos API goes dark on September 24, 2026, and the fix isn't a version bump — it's a full swap to a different vendor, a different pricing model, and in some cases a different way of calling the API entirely. If you've already read the shutdown notice, this is the part it didn't cover: what to do about it before Thursday.

What Exactly Shuts Down on Thursday

OpenAI's own Sora API documentation now opens with a banner: the Sora 2 video generation models and the Videos API "are deprecated and will shut down on September 24, 2026." That covers the API endpoints and five model identifiers — sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06 — not just a consumer feature. The consumer Sora app and web experience already shut down earlier this year; OpenAI's own Sora landing page still carries the line "As of April 26, 2026, the Sora product is no longer available." The API was the last piece running, and it stops in days, not months.

One wrinkle worth flagging before you copy any values into code: OpenAI's guide states plainly that "both sora-2 and sora-2-pro support 16- and 20-second generations," and its own batch-API example requests exactly that — a 20-second sora-2-pro render at 1920x1080 and a 16-second one at 1080x1920. But the separate API reference page for the create-video endpoint still lists the allowed seconds values as only 4, 8 or 12, defaulting to 4. The two OpenAI pages disagree; take the guide's longer durations as the current behavior and don't be surprised if the reference catches up.

What Sora's Videos API Actually Did

Worth knowing exactly what you're losing, because the replacement has to cover the same ground. Per OpenAI's guide, sora-2 was built for speed — fast, cheap iteration, "often more than sufficient for social media content, prototypes." sora-2-pro was the production tier, the only one that reached 1080p in 1920x1080 or 1080x1920. Both models could take an image reference to guide a generation, reuse character assets across renders, extend or edit an existing clip, and run through the Batch API for large offline render queues — a JSON-only path that lets you queue dozens of shots at once and download completed videos for up to 24 hours after the batch finishes. The call itself was asynchronous: POST /videos returns a job ID and a queued status, you poll GET /videos/{id} or register a webhook, and once the job completes you pull the MP4 from GET /videos/{id}/content. The same endpoint could also hand back a thumbnail or a spritesheet for scrubbers and previews, using a variant query parameter.

What OpenAI Says About Your Existing Work

OpenAI's own help center article on the discontinuation is specific and worth acting on today, not next week: export your content now at sora.chatgpt.com/sunset, and expect an email when the export is ready. "After Sora is discontinued, and after the period of time of any final export window passes (if we are able to offer one), we will permanently delete any data associated with your use of Sora," the page states. Any purchased ChatGPT or Sora credits carry over to Codex use, but not to another video product. There's no stated grace period for API-generated assets beyond the same export mechanism; if you've been calling the Videos API and letting finished clips sit in OpenAI's storage instead of downloading them, treat this week as the deadline to pull everything down.

Where Each Sora Job Goes Now

A quick social clip, the job Sora 2 itself was tuned for, has the cheapest home on Google's Gemini API: Veo 3.1 Lite (veo-3.1-lite-generate-preview) runs $0.05 per second at 720p and $0.08 at 1080p, per Google's own pricing page. On fal, MiniMax H3 Max undercuts that further at the fastest tier — $0.025 per second at 480p, $0.04 at 768p — with up to 15-second clips and native audio.

A 1080p export, sora-2-pro's job, has three real options. Veo 3.1 Standard (veo-3.1-generate-preview) charges $0.40 per second at both 720p and 1080p, $0.60 at 4K. Alibaba Cloud Model Studio's Wan 3.0 (wan3.0-video) is cheaper at that resolution — $0.20 per second at 1080p — and, per Alibaba's own Model Studio page, renders a single continuous take up to 30 seconds, ten seconds longer than Sora ever went. MiniMax H3 Max at 1080p runs $0.16 per second on fal's standard tier or $0.08 on its Turbo tier.

An image-to-video shot, Sora's input_reference feature, maps most directly onto Kling 2.6 Pro on fal, which generates from an image with native audio at $0.07 per second with audio off or $0.14 with it on — though its outputs cap at 5 or 10 seconds, shorter than Sora's grid. Black Forest Labs' FLUX 3 (flux-3-video) also does image-to-video on one endpoint, priced from $0.17 per second for a full HD render or $0.06 for a draft preview, per BFL's own pricing docs and knowledge base.

A batched programmatic job, the shot-list workflow Sora ran through its Batch API, has a close cousin in ByteDance's Seedance 2.5 on fal, which per fal's own explainer accepts up to 50 reference files in one request — stills, clips and audio combined — and renders a single continuous 30-second take. It's the one model here that doesn't bill per second at all.

The Calling-Shape Trap: Tokens vs. Seconds

This is where a working Sora pipeline is most likely to break silently. Seedance 2.5 bills by token, not by second: $0.0214 per 1,000 tokens, where tokens scale with resolution, duration and frame rate — about $0.4730 per second of finished 720p video by fal's own math, but a number your code has to compute rather than read off a price sheet. Everything else here — Veo 3.1, Wan 3.0, MiniMax H3, Kling, FLUX 3 — bills by the second, matching Sora's model more closely. According to WaveSpeedAI's own migration writeup, Sora's fixed duration grid of 4, 8, 12, 16 and 20 seconds doesn't exist on any of these platforms; each has its own duration ceiling and increments, so a size-and-seconds table built for Sora needs rebuilding, not just swapping in a new base URL. The async job-plus-poll pattern does carry over cleanly to most of these APIs, which is the one part of a Sora integration you likely won't have to touch.

What to Test Before You Commit

Before pointing production traffic anywhere, check five things against the new vendor's own docs rather than assuming Sora parity: audio support (Google's Veo 3.1 and BFL's FLUX 3 both confirm native, synchronized audio on their own pages; verify it explicitly for anything else you pick), aspect ratios and duration ceilings (several of these cap well under Sora's 20 seconds), reference-image or reference-file limits (Seedance 2.5's 50-file cap is generous, Kling 2.6 Pro takes a single image), commercial licence terms (BFL publishes per-model licences for its FLUX line; check the equivalent page for any model before shipping client work), and rate limits under real load, which none of these pricing pages state and which only shows up once you're running production volume.

What to Watch in the First Week

Expect two failure modes once you've switched: cost surprises from token-billed models where a duration or resolution change moves the price non-linearly, and prompt drift, since none of these models were trained on Sora's prompting conventions and a prompt tuned for sora-2 won't behave identically on Veo, Wan or Seedance. Budget a real week of side-by-side testing against your old Sora outputs before the old outputs themselves become impossible to regenerate for comparison.

Every model above takes an API key and a prompt today. OpenAI's clock is the only one that's fixed — it runs out September 24.