The short answer: which video models replace Sora
Sora’s API stops working on September 24, 2026. The models you would move to are Veo 3.1, Kling V3, Seedance 2.0 and 2.5, Wan 2.7 and 3.0, Hailuo 2.3, Vidu Q3, Grok Imagine, SkyReels and PixVerse — and they cost anywhere from $0.035 to $0.569 per second depending on resolution, audio and length. The useful question is not which one is best; it is which constraint you cannot move, because the models differ most on three things: how long a clip they will produce, how high they will render at that length, and whether audio is native or an option you pay for. This page gives you the per-second price of each, plus the request fields you have to change when you leave Sora’s API behind.
This article is from Modellix, an AI model API aggregator — we sell metered access to most of the models below, so treat the price tables as commercially interested, and check them against each vendor yourself. We do not carry Sora and never have, which is why there is no “stay with us” pitch at the end of this comparison. One audience note before we start: if what you actually want is a finished editing app you open in a browser, this page will not help — everything here is an API you call from code.
Sora’s API shuts down on September 24, 2026 — the timeline that matters
Most pages ranking for Sora alternatives still describe a live product. It is not live. OpenAI’s own documentation, starting with its Sora discontinuation notice, states the sequence plainly:
- April 26, 2026 — “The Sora web and app experiences were discontinued on April 26, 2026.” OpenAI’s Sora 2 product page now opens with the same date.
- September 24, 2026 — “The Sora API will be discontinued on September 24, 2026.” The deprecation notice names the affected surface: the Videos API and the models
sora-2,sora-2-pro,sora-2-2025-10-06,sora-2-2025-12-08, andsora-2-pro-2025-10-06. - After the export window — OpenAI says that once the final export window closes — if it is able to offer one — it will “permanently delete any data associated with your use of Sora”, and recommends exporting soon. The export lives at
sora.chatgpt.com/sunset.
Three details are worth pulling out before you plan the migration. First, the deprecation notice in OpenAI’s video generation guide still documents the API in full, so anything you build against it works until the cutoff and then stops — there is no gradual degradation. Second, the price you are leaving is high at the top end: OpenAI’s published rates are $0.10 per second for sora-2 at 720p, and $0.30 / $0.50 / $0.70 per second for sora-2-pro at 720p, 1024p and 1080p, halved on the Batch tier. Third, the duration envelope was genuinely long: both models generate 16- and 20-second clips, and extensions can add up to 20 seconds each up to six times, for a maximum of 120 seconds.
If you were using Sora for its longest, highest-resolution jobs, that combination — 20 seconds at 1080p — is the single hardest thing to replace on this list, and it is the first thing to check in the table below. For the endpoint-level view of the API you are leaving, our notes on OpenAI’s video generation API cover what still works and what the deprecation covers.
What you actually have to change when you leave Sora
The migration is smaller than it looks, because the request shapes are cousins. Sora’s Videos API takes four things that matter: model, prompt, size (for example 1280x720 or 1920x1080) and seconds — and the docs show seconds as a string, like "8" or "20". Everything else — polling, task states (queued, in_progress, completed, failed), the option to hand the job off rather than poll — is scaffolding you already have. If you want the short version of this section first, start with the five-point migration checklist.
How the four fields you send today map onto the replacement APIs. Field names and ranges were read from each provider’s live request schema on September 15, 2026.
The differences that will actually cost you an afternoon:
The model becomes part of the URL. Instead of a model field, most aggregator and vendor endpoints take the model as a path segment — on Modellix, POST /api/v1/<provider>/<model>. Your “which model am I calling” logic moves out of the body and into configuration, which is usually an improvement.
size splits in two. Replacement APIs want a resolution tier (720p, 1080p, 4k) plus an aspect ratio — and the aspect-ratio field name is not standardised. Veo 3.1 calls it aspectRatio; Kling V3, Grok Imagine, Vidu Q3 and PixVerse call it aspect_ratio; Seedance and Wan call it ratio. If you are writing an abstraction layer, this is the field that forces you to keep a per-model parameter map rather than a single struct.
seconds becomes an integer with a per-model range. Sora took a string from a small set. Replacements take an integer, and the legal range belongs to the model: 4–8 on Veo 3.1 (as values 4, 6, 8), 3–15 on Kling V3, 4–15 on Seedance 2.0, 2–30 on Wan 3.0 and 4–30 on Seedance 2.5. Two of these bite immediately: on Veo 3.1, 1080p and 4k are only available at an 8-second duration, so “high resolution” and “short clip” are mutually exclusive; and on Hailuo 2.3, 1080p is capped at 6 seconds, so asking for 10 seconds silently forces 768p.
Audio changes from included to a switch. Sora 2 shipped as a video-and-audio model. Some replacements keep audio on by default (Seedance’s generate_audio defaults to true; Wan 3.0’s audio defaults to true), some default to silence and charge more when you enable it (Kling V3’s audio defaults to off, and a 1080p clip costs $0.0896/s with audio off versus $0.1344/s with native audio), and some have no audio parameter at all. Field names differ again: generate_audio, sound, generate_audio_switch.
Character consistency has no single drop-in. Sora let you upload two- to four-second clips to register a character, and accepted up to two characters per video. Nothing on this list matches that field-for-field. The closest options are Kling V3 Omni, which takes a reference video; Wan 3.0 I2V, which fuses up to ten reference images; and Seedance 2.0 I2V, which accepts mixed image and audio references.
Video API Migration Reference
See the submit, poll and webhook fields for each video model, with per-model request schemas.
View DocsWhat video costs per second on every serious alternative
Prices below were read from each model’s live page on September 15, 2026, at the tier most people actually deliver at — 1080p where the model offers it, otherwise its top tier. Rates change; re-check before you commit a budget. Vendor list rates run above the platform rates below rather than matching them: Google’s Gemini API pricing publishes Veo 3.1 with audio at $0.40 (720p and 1080p) and $0.60 (4k), while the equivalent platform tier here is $0.36 and $0.54.
| Model | Duration range | Resolution tiers | Audio | Price per second (USD) |
|---|---|---|---|---|
| Vidu Q3 Turbo | 1–16s | 540p / 720p / 1080p | not exposed | $0.035 – $0.065 |
| Seedance 2.0 Mini | 4–15s | 480p / 720p | on by default | $0.04 – $0.08 |
| Veo 3.1 Lite (Google) | 4/6/8s | 720p / 1080p | native | $0.045 – $0.072 |
| Vidu Q3 Pro | 1–16s | 540p / 720p / 1080p | not exposed | $0.045 – $0.12 |
| Wan 3.0 (Alibaba) | 2–30s | 480P / 720P / 1080P | on by default | $0.05 – $0.20 |
| Hailuo 2.3 (MiniMax) | 6s or 10s | 768P / 1080P | not exposed | $0.0504 – $0.0738 |
| Seedance 2.0 Fast | 4–15s | 480p / 720p | on by default | $0.06 – $0.12 |
| PixVerse C1 | 1–15s | 360p / 540p / 720p / 1080p | option | $0.06 – $0.24 |
| Kling V3 (audio off) | 3–15s | 720p / 1080p / 4k | option | $0.0672 – $0.336 |
| Seedance 2.0 | 4–15s | 480p / 720p / 1080p / 4k | on by default | $0.07 – $0.78 |
| Grok Imagine (xAI) | 1–15s | 480p / 720p | not exposed | $0.084 – $0.108 |
| Veo 3.1 Fast | 4/6/8s | 720p / 1080p / 4k | native | $0.09 – $0.27 |
| Wan 2.7 | 2–15s | 720P / 1080P | audio input only | $0.09 – $0.135 |
| SkyReels T2V | 3–15s | 480p / 720p / 1080p | option | $0.096 – $0.42 |
| Kling V3 (native audio) | 3–15s | 720p / 1080p / 4k | option | $0.1008 – $0.336 |
| Seedance 2.5 | 4–30s | 480p / 720p / 1080p | on by default | $0.103 – $0.569 |
| Veo 3.1 (Google) | 4/6/8s | 720p / 1080p / 4k | native | $0.36 – $0.54 |
Rows are ordered by their lowest tier. For reference: Sora’s own rates were $0.10/s (sora-2, 720p) and $0.30–$0.70/s (sora-2-pro, 720p–1080p). Prices are platform list rates as of September 15, 2026 and exclude account-level top-up discounts.
Each bar is the tier you would deliver at — 1080p where the model offers it, otherwise 720p — with Kling V3’s 4k (audio off) tier kept in as well, because that is where its highest rate sits. The two grey bars are Sora’s own published rates, kept in the same chart so the comparison is like-for-like.
The spread is 16× at the extremes — Vidu Q3 Turbo’s $0.035/s at 540p against Seedance 2.5’s $0.569/s at 1080p; across the delivery tiers shown in the chart, the range is $0.065 to $0.569. The cheapest option is not the cheapest by accident, as the model notes below explain. If you want the same field read from a different angle, our video API shortlist with per-second prices benchmarks the same models on real clips.
The alternatives, model by model
Eight entries follow, in no ranking order — each one leads with the constraint it is best at, and each price is the tier you would actually ship. If you would rather see the whole field ranked by quality with sample clips, our roundup of the best AI video generators covers that angle; this page stays on migration cost and field-level fit.
Veo 3.1 (Google)
The closest thing to a Sora successor in kind: multilingual, cinematic, native audio, and available in three price tiers so you can trade quality for rate. Standard 3.1 is the premium pick at $0.36/s (1080p) and $0.54/s (4k); Veo 3.1 Fast lands at $0.09–$0.108/s for 720p–1080p; Veo 3.1 Lite is the budget tier at $0.045–$0.072/s with no 4k at all. The constraint to plan around: duration choices are 4, 6 or 8 seconds, and 1080p and 4k require the full 8. If Sora’s 20-second clips were load-bearing for you, Veo is a step down in length and a step up in per-second cost. Our Veo 3.1 price breakdown walks the tiers.
Kling V3 (Kuaishou)
The best match for Sora’s long-take ambition without the premium: 3–15 seconds as an integer, up to 4k, with multi-shot planning on by default and audio as a paid switch. Pick audio off and 1080p costs $0.0896/s; turn native audio on and it is $0.1344/s. Two honest caveats: there is no separate negative-prompt field (the model takes positive and negative descriptions in one prompt), and the 4k tier prices at $0.336/s whether or not you want audio.
Seedance 2.0 and 2.5 (ByteDance)
Seedance is where the price range is widest and the trade-off is clearest. Seedance 2.0 spans 4–15 seconds at 480p through 4k, with generate_audio, seed, camera_fixed and return_last_frame all exposed as parameters — but the T2V endpoint is text-only and rejects image, video and audio inputs, and 4k costs $0.78/s. Its siblings are the value picks: Seedance 2.0 Fast at $0.06–$0.12/s and Mini at $0.04–$0.08/s, both capped at 720p. Seedance 2.5 buys the one thing Sora had and Veo cannot match — clips up to 30 seconds — at $0.103/s (480p) to $0.569/s (1080p).
Wan 2.7 and 3.0 (Alibaba)
The cheapest way to get a long clip at 1080P: Wan 3.0 runs 2–30 seconds at 480P/720P/1080P with audio on by default, priced at $0.05/s (480P), $0.10/s (720P) and $0.20/s (1080P). The older Wan 2.7 is 2–15 seconds at 720P/1080P for $0.09–$0.135/s, and its audio field is an input (audio_url, 2–30 seconds, up to 15MB) rather than a generation switch — useful if you already have a voice track.
Hailuo 2.3 (MiniMax)
Hailuo is the character-performance specialist: faces, emotion, dialogue scenes, stylised art, at $0.0504–$0.0738/s, which makes it one of the cheapest credible 1080P options. The catch is structural rather than occasional: duration accepts only 6 or 10 seconds, and 1080P is limited to the 6-second version — 10 seconds forces 768P. If your clips are short and human-centric, it is the best value on the page; if you need 1080p for ten seconds in one shot, it cannot do it.
Vidu Q3 (Shengshu)
The cheapest general-purpose video on the list: Vidu Q3 Turbo at $0.035/s (540p), $0.055/s (720p) and $0.065/s (1080p), with Pro at $0.045–$0.12/s for more cinematic output. Both accept 1–16 seconds — a longer window than Veo or Kling — and a seed for reproducibility. Neither exposes an audio parameter, so plan to add sound downstream.
Grok Imagine (xAI)
Short, social-first clips: 1–15 seconds, seven aspect ratios, but a hard ceiling of 480p/720p at $0.084–$0.108/s. One thing worth stating plainly, because it cuts against the usual pitch: these are aggregator rates, and they sit above xAI’s own published list price — so if Grok Imagine is the only model you need, going direct is cheaper. It earns its place here if you are consolidating several models behind one key.
SkyReels and PixVerse C1 (the cheap end, with strings)
SkyReels T2V covers 3–15 seconds at 480p–1080p for $0.096–$0.42/s, but its fast mode cannot generate audio (sound must be false). PixVerse C1 is the most granular on resolution — 360p, 540p, 720p, 1080p — at $0.06–$0.24/s, with duration, aspect_ratio and quality all required rather than defaulted, and no multi-clip support. Both are reasonable draft-tier choices and poor choices for a final delivery at 1080p.
Two names you will see on every other Sora-alternative list are Luma Dream Machine and Runway, and neither is in our catalog, so we cannot quote their per-second cost — Runway in particular meters in credits rather than dollars per second, which is exactly why it does not appear in the table above. Check their pricing pages directly if either is on your shortlist.
One thing a price table cannot tell you: watermarking and provenance metadata policy differ by provider and sometimes by tier, and none of the vendors above standardise on one scheme. If you ship client work — or anything an audience may republish — verify the current policy on the model page you pick rather than assuming it matches what you had on Sora.
How to choose: start from the constraint you cannot move
Don’t rank these models; filter them. Find the one row that describes your job and start there — the list narrows fast.
| If your hard constraint is… | Start with | Because |
|---|---|---|
| Clips longer than 15 seconds | Seedance 2.5 or Wan 3.0 | Both reach 30 seconds; Wan is the cheaper of the two |
| 1080p at the lowest cost | Vidu Q3 Turbo, then Hailuo 2.3 | $0.065/s and $0.0504–$0.0738/s respectively |
| Highest fidelity, budget secondary | Veo 3.1 or Kling V3 | $0.36/s and $0.336/s at the top tiers |
| Native audio baked into the clip | Veo 3.1, Seedance 2.0/2.5, Wan 3.0 | Audio is on by default rather than an upsell |
| Character or product consistency | Wan 3.0 I2V, Kling V3 Omni, Seedance 2.0 I2V | Reference images or reference video instead of Sora’s character upload |
| A draft tier you can batch cheaply | Seedance 2.0 Mini, Vidu Q3 Turbo | $0.04–$0.08/s and $0.035/s at their low tiers |
If two of these apply, you are the reader who should run both for a week rather than pick now. For a head-to-head on the pair that most often collide on price and audio, see Seedance vs Kling.
What the switch does to your bill
Do the arithmetic on your own volume before you commit, because the per-second number hides the multiplier. A 20-second clip on sora-2 at 720p cost $2.00 in generation. On Wan 3.0 the same 20 seconds at 480P costs $1.00 and at 1080P costs $4.00; on Seedance 2.5 at 1080p it costs $11.38. Going the other way, an 8-second 1080p clip that used to cost $5.60 on sora-2-pro costs $2.88 on Veo 3.1 — just over half, for a clip two and a half times shorter than Sora’s maximum.
Two accounting details matter as much as the rate. You pay for output duration, not wall-clock time or retries you cancel before completion — so a generous resolution tier used for testing gets expensive quietly. And Sora’s Batch tier halved its rates, so if your workload was batch-friendly, the honest baseline to beat is $0.05–$0.35 per second, not the list price. Our cheapest image and video API routes covers the low end of that comparison.
Running several of them behind one key
The reason to keep more than one of these in your stack is that the constraint table is model-specific, and it changes: a model that caps at 720p today ships 1080p next quarter. Splitting your pipeline across providers keeps you portable, but it also means four keys, four invoices and four failure modes — which is where an aggregation layer earns its keep rather than pretending to be the cheapest option for any single model.
Modellix’s media API takes the model in the path and settles per second of output at the tier you requested — 178 image, video and speech models carried published rates on September 15, 2026. It accepts an X-Webhook-URL header for terminal-state callbacks and keeps uploaded inputs about seven days, free, for reuse as references; the REST API reference documents both, and per-request cost lands in the request logs, so you can reconcile the tables above against your invoice. The mechanics are in routing several video models through one API key.
Compare Sora replacements on one key
Log in to run the same prompt across Veo, Kling, Seedance and Wan and compare per-second cost.
LoginFrequently asked questions about Sora alternatives
Is the Sora API definitely shutting down?
Yes, as of September 15, 2026. OpenAI’s documentation states the API will be discontinued on September 24, 2026, and the consumer product closed on April 26, 2026. The API still works until the cutoff, so anything built on it runs normally and then stops.
Is there a free Sora alternative?
Not on the API side. The alternatives above bill per second of output with no permanent free tier, and the free access most people remember was in Sora’s consumer app, which no longer exists. Google’s AI Studio lets you experiment with Veo before you commit, and the cheapest tiers here (Vidu Q3 Turbo at $0.035/s, Seedance 2.0 Mini at $0.04–$0.08/s) are inexpensive enough to prototype with, but they are not free.
What happens to the content I already made in Sora?
OpenAI says data associated with your use of Sora will be permanently deleted after any final export window closes, and recommends exporting as soon as possible. The export tool is at sora.chatgpt.com/sunset. Treat this as a deadline separate from the API cutoff.
Which alternative is closest to Sora 2 overall?
There is no one-to-one. Veo 3.1 is closest on cinematic quality with native audio, Kling V3 is closest on length and control, and Seedance 2.5 is the only family here that exceeds Sora’s 20-second maximum, reaching 30 seconds. If you depended on Sora’s character uploads specifically, expect to rebuild that step around reference images or a reference video.
Can I keep using Sora until September 24?
Yes. The API is deprecated, not disabled, and OpenAI’s guide still documents the request format. Building anything new on it is a poor use of the time you have left.
Do I have to rewrite my whole pipeline to switch?
No. The task lifecycle — submit, poll, retrieve — is the same. What changes is the field set: the model moves into the URL, size becomes a resolution tier plus an aspect-ratio field whose name varies by provider, and seconds becomes an integer duration with a per-model range. Most teams are looking at a parameter map and a new base URL, not a rearchitecture.
Provider details and pricing reflect public information as of September 15, 2026, and change frequently — especially around the Sora shutdown. Validate against each provider’s live pricing before committing. Access image and video models, including Veo 3.1, Kling V3 and Seedance 2.0, through a single API key at modellix.ai.
