Sora 2 is shutting down: alternatives and migration guide

OpenAI's Sora 2 API (the sora-2 and sora-2-pro models, plus the Videos API that serves them) sunsets on September 24, 2026, 58 days from this guide's publish date. If your pipeline calls those endpoints, the closest replacements are Seedance 2.0 Pro for native audio-plus-video output and Kling 3.0 Pro for camera-driven cinematic shots, both available now through the same API pattern Sora 2 used.

OpenAI notified developers of the deprecation on March 24, 2026. The consumer Sora app and web experience had already closed on April 26, 2026, so the API was the last piece still running (OpenAI's Sora discontinuation notice). The date is confirmed independently on OpenAI's own deprecations page, which lists sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08, and sora-2-pro-2025-10-06 all retiring on the same day, and again by third-party tracking at apiyi.com. Reporting has attributed the shutdown to unsustainable compute cost, with figures circulating around $1 million a day in burn, alongside a monthly active user decline from a peak near 1 million down to under 500,000. Those are numbers from press coverage, not figures OpenAI itself has published (80.lv). No Sora 3 has been announced to replace it.

Infer hosts all of these models, so we have no favorite. The mapping below is built from spec sheets and current Infer leaderboard positions, not a sales pitch for any one of them.

The Sora 2 timeline, in order

Three separate dates get conflated in most coverage of this shutdown, and they matter for different reasons depending on what you built:

• March 24, 2026: OpenAI notifies developers via the API deprecations page that sora-2, sora-2-pro, and the Videos API are scheduled for removal.
• April 26, 2026: the consumer Sora app and sora.com web experience close. If you only ever used the app, this is the date that already affected you.
• September 24, 2026: the API itself stops serving requests. This is the date that matters for anything programmatic: scripts, product features, automated content pipelines.

If your integration is still live today, you're operating in the six-month gap between the second and third date. That's the window this guide is written for.

Why not Wan, Hailuo, or HappyHorse

A fair mapping should say why three models made the cut and not others. Wan 2.2 T2V-A14B is open weights and cost-efficient, but Infer's own comparative claim puts it at roughly 85% of Hailuo 02 Pro's quality, and it has no confirmed native audio, which rules it out as a Sora 2 audio replacement. Hailuo 02 Pro is the cheapest of the bunch at $0.08/sec, but its resolution, duration, and audio specs aren't independently confirmed on Infer's own model page (that page has been returning a server error during our data collection, with only meta-tag text recoverable), so we're not routing a "closest match" recommendation through unverified specs. HappyHorse-1.0 tops Infer's own leaderboard at 1368 Elo, but per Infer's own leaderboard notes it has no public API yet, which disqualifies it outright for a migration guide about API replacements.

What replaces what

Sora 2's headline feature was native audio generated in the same pass as the video, which narrows the field to two real candidates on architecture alone. Kling gets a slot too: teams that used Sora 2 mainly for camera work and shot composition, not audio, will find it the closer visual match.

Elo figures are from the Artificial Analysis Video Arena snapshot Infer's leaderboard cites, dated 2026-04-30. Artificial Analysis's live arena has since reshuffled these same three models to different ranks, so treat the numbers above as a dated snapshot, not a permanent ordering (Infer leaderboards).

We're skipping a side-by-side of old Sora 2 pricing against these three. OpenAI's Sora 2 API rate card isn't independently confirmed in our source pack, and guessing at it would violate the rule that matters more than completeness here: never publish a number you can't source.

What a 10-second clip costs after migration

Take a 10-second clip, the length most Sora 2 demo reels used. On Infer, that's $1.30 with Seedance 2.0 Pro, $0.90–$1.00 with Veo 3.1 Fast, and $1.00 with Kling 3.0 Pro, all billed only for generations that succeed. Run that same clip length across a 50-clip test batch while you're validating the migration and the spread becomes real money: $65 on Seedance versus $45–$50 on Veo, which is worth knowing before you pick a default model for high-volume testing rather than after. Full worked examples for larger jobs are in the AI video pricing index.

Migration steps

1. Swap the request pattern, not the whole architecture. Sora 2's Videos API used an async job-creation-then-poll pattern: submit a generation request, get a job ID back immediately, poll for status, then retrieve the finished asset. All three replacements work the same way on Infer. Seedance, Veo 3.1 Fast, and Kling 3.0 Pro are async/queued endpoints that return a job ID on submission, with a 60 requests/minute default rate limit. The plumbing you already built for polling and retries carries over; only the endpoint URL, auth header, and response schema change.

2. Rewrite prompts in three parts, not one paragraph. Sora 2 prompting rewarded a single descriptive paragraph with physics and lighting folded in. Seedance and Kling both respond more predictably to prompts split into subject/action, camera movement, and style as separate clauses. Run a batch of your highest-value existing prompts through both structures and compare before committing to one house style.

3. Test audio separately from video quality. If you're moving to Seedance 2.0 Pro for the audio parity, don't assume Foley and dialogue transfer cleanly from a Sora 2 prompt written before audio was a factor in your review process. Seedance's audio is generated jointly with the video rather than added after, which changes how word choice in the prompt affects the soundtrack. Test with the same prompt at least twice before locking a template.

4. Re-check resolution and duration ceilings against your current outputs. Kling 3.0 Pro ships at 1080p from a native 4K capture pipeline. Veo 3.1 Fast tops out at 1080p natively but chains 8-second segments up to roughly 148 seconds. Seedance 2.0 Pro runs 720p at 5–10 second clips. If any current Sora 2 workflow depends on a duration or resolution these don't hit natively, flag it now, before discovering it in production on September 25.

5. Budget the cutover as a project, not a code diff. Set a hard internal deadline well before September 24, 2026. A week of buffer for unexpected API quirks is cheap insurance against a hard stop with zero fallback.

Testing checklist before the cutoff

Run through this before September 24, 2026, not the week of:

• Re-run your five highest-traffic prompts on the replacement model and compare output against your last known-good Sora 2 render, side by side.
• Confirm your retry and error-handling logic accounts for the new provider's job-failure response format; it won't match Sora 2's error codes exactly.
• Check whether any downstream process rejects or flags watermarked output; Veo 3.1 Fast embeds a visible watermark plus SynthID-Audio on every generation, which matters if a client-facing deliverable can't carry one.
• Load-test at your real request volume against the 60 requests/minute default rate limit; ask Infer support about limit increases if your volume is higher.
• Set a calendar reminder for two weeks before the cutoff to run a full end-to-end pass, beyond whatever unit tests already cover the integration.

The verdict

Choose Seedance 2.0 Pro if audio-plus-video in a single generation is the feature you can't lose. It's the architecturally closest thing to what Sora 2 did, at $0.13 per second. Choose Veo 3.1 Fast if you need long-form output stitched from shorter chained clips and can live with its visible watermark. Choose Kling 3.0 Pro if audio was never the point and camera work and shot consistency are what your Sora 2 pipeline actually leaned on. If your workload genuinely needs all three (audio parity, long duration, and camera control), the honest answer is that no single model replaces Sora 2 one-for-one, and you should pick per-project rather than forcing one model to do everything Sora 2 did.

Related reading: the best AI video models in 2026, ranked, the cheapest AI video generation APIs, the state of AI video generation, July 2026, the AI video pricing index, and the guides hub for every how-to and migration guide. All three replacement models above run through a single Infer API key: try Seedance 2.0 Pro on Infer →, try Veo 3.1 Fast on Infer →, try Kling 3.0 Pro on Infer →.