Sora 2: What It Was, Why It Ended, and What to Use Instead (Updated August 2026)
Last updated August 7, 2026
Sora 2 was OpenAI's video model with native synced audio and physics-aware realism, launched September 30, 2025. OpenAI shut down the Sora app and sora.com on April 26, 2026, and the API is scheduled to end September 24, 2026. Exported videos remain yours; for new generations, current models like Veo 3.1, Kling 3, Seedance 2.0, and MiniMax-H3 cover Sora 2's jobs.
Updated August 2026
Sora 2 brought natively synchronized audio and physics-aware realism to mainstream attention when it launched on September 30, 2025 — and less than seven months later, OpenAI was shutting it down. It was the company's flagship AI video model; as of August 2026, it is a sunsetting product: OpenAI discontinued the Sora app and sora.com on April 26, 2026, and the Sora API is scheduled to shut down on September 24, 2026. Videos already exported remain yours to keep and use; the ability to generate new ones through OpenAI is ending. This page is the dated record of what Sora 2 actually was, what its real limits were (including the widely misreported maximum length), and which current models cover the jobs it used to do.
What happened to Sora? The full timeline
Most coverage of Sora is frozen at some point in its lifecycle — either the 2024 research-preview hype or the Sora 2 launch. Here is the complete arc, with dates.
| Date | Event |
|---|---|
| Feb 15, 2024 | Sora unveiled as a research preview — a diffusion-transformer text-to-video model generating up to 60-second clips, framed as "video generation models as world simulators." Never publicly shipped in this form. |
| Dec 9, 2024 | Sora Turbo launches at sora.com as a ChatGPT Plus/Pro perk — up to 1080p and 20 seconds, with Storyboard, Remix, Blend, and Loop tools. No audio. |
| Sept 30, 2025 | Sora 2 and Sora 2 Pro launch alongside an invite-only iOS app (US/Canada) with a short-video feed and Cameos. OpenAI calls it a possible "GPT-3.5 moment for video." |
| Oct 6, 2025 | sora-2 and sora-2-pro arrive in the OpenAI API at DevDay. |
| Nov 2025 | Android app expands access to Japan, South Korea, Taiwan, Thailand, and Vietnam; OpenAI calls the free tier "completely unsustainable" and introduces paid extra generations — just 32 days after launch. |
| Dec 11, 2025 | Disney invests $1B in OpenAI and licenses roughly 200 Disney, Pixar, Marvel, and Star Wars characters as cameos, with a stated one-year Sora exclusivity. |
| Mar 13, 2026 | Sora 1 is sunset for US users and Sora 1 libraries are permanently deleted — in hindsight, the first public signal of the wind-down. |
| Mar 24, 2026 | OpenAI announces the discontinuation via an X post and a help-center FAQ — no blog post. Disney exits the partnership. |
| Apr 26, 2026 | The Sora app and sora.com shut down; unexported user data is deleted after the export deadline. |
| Sept 24, 2026 | Scheduled shutdown of the Sora API (still functional as of August 2026). |
The dates carry the story: 32 days from launch to "completely unsustainable," roughly fifteen weeks from Disney's $1B investment to the discontinuation announcement. OpenAI has said Sora continues internally as a research program on world models rather than a consumer product.
What were Sora 2's actual specs?
Sora 2's specs are among the most misquoted on the internet, largely because the app and the API had different limits, and the API quietly gained capabilities that were only documented in OpenAI's developer cookbook.
| Spec | Sora 2 (app / sora.com) | Sora 2 / Pro (API) |
|---|---|---|
| Resolution | Up to 1080p (Pro) | sora-2: 720x1280 / 1280x720; sora-2-pro adds 1024x1792 / 1792x1024 and 1080x1920 / 1920x1080 |
| Clip length | 10s at launch, later 15s; 25s for Pro on web via storyboards | 4, 8, 12, 16, or 20 seconds per generation |
| Maximum continuous video | 25s (Pro, web) | 120 seconds, via up to 6 chained extensions |
| Audio | Native synchronized dialogue, sound effects, ambience | Same |
| Inputs | Text-to-video, image-to-video, remix, Cameos, licensed character cameos | Text-to-video, image-to-video (image anchors the first frame), extensions, Characters API |
| Frame rate | Never officially specified (observers measured ~30fps at 720p) | Never officially specified |
| API pricing | — | sora-2: $0.10/s; sora-2-pro: up to $0.70/s at 1080p (per OpenAI's model page) |
Correcting the record on length: the claim that "Sora 2 maxes out at 25 seconds" appears in most explainers and is wrong for the API. OpenAI's Sora 2 prompting guide documents 16- and 20-second generations plus extension chaining to a true 120-second maximum — a capability that outlived the consumer app by five months. The API also shipped a Characters API (POST /v1/videos/characters) that built reusable characters from 2–4 second reference clips, with human likenesses gated behind account-manager approval.
What made Sora 2 different from other video models?
Two things, both of which reset expectations for the whole category.
Physics that fails honestly. OpenAI's launch post emphasized that Sora 2 modeled failure, not just success: a missed basketball shot "will rebound off the backboard" rather than teleporting into the hoop, and it could handle "backflips on a paddleboard that accurately model the dynamics of buoyancy and rigidity." Earlier video models tended to warp reality to satisfy the prompt; Sora 2 let physical outcomes play out.
One-pass synchronized audio. Sora 2 generated dialogue, sound effects, and ambient sound together with the picture — the headline change from every prior Sora, and the feature that made single-generation clips feel like finished footage rather than silent b-roll.
The model also had genuine creative champions. Filmmaker Paul Trillo, who directed Washed Out's "The Hardest Part" — billed as the first commissioned Sora music video — said "Working with Sora is the first time I've felt unchained as a filmmaker." And the Cameos feature — inserting a verified, consenting likeness of yourself into generated scenes — became the app's true product-market fit, most famously through the wave of Sam Altman cameo memes in October 2025.
Why did OpenAI shut down Sora?
Economics, primarily. Reporting sourced to the Wall Street Journal put Sora's compute costs at roughly $1 million per day while usage collapsed — from around one million users at peak to under 500,000 — and the app fell from #1 on the App Store to #165 by March 2026. Those figures are journalists' estimates, not OpenAI disclosures, but OpenAI's own actions corroborate the direction: it declared the free tier "completely unsustainable" barely a month after launch and began charging for extra generations in November 2025.
The strategic backdrop was a pivot toward coding, enterprise, and agent products. The Disney partnership — announced with a one-year exclusivity in December 2025 — ended roughly fifteen weeks later when the shutdown was announced. Notably, OpenAI never published a blog post about the discontinuation; the news went out via an X post and a help-center FAQ.
Did Sora videos have watermarks?
Yes — but there were two different regimes, and almost every explainer conflates them:
- Consumer (app and sora.com): every download carried a visible moving watermark plus C2PA Content Credentials metadata. OpenAI also ran internal reverse video/audio search to trace outputs.
- API: outputs carried C2PA metadata only — no visible watermark.
Watermark-removal tools circulated within a week of the Sora 2 launch (reported by 404 Media in October 2025), which is why C2PA metadata, not the visible mark, was the more durable provenance layer.
Which Sora prompting lessons still matter?
OpenAI's own Sora 2 prompting guide remains one of the best short documents ever written on directing a video model, and its advice transfers directly to current models. Its core lessons, credited to that guide:
- Treat the model like "a cinematographer who has never seen your storyboard." Spell out everything; re-running the same prompt yields different results by design.
- One camera move and one subject action per shot. Overloaded shots produce mush on every model, not just Sora.
- Describe action in countable beats — "cyclist pedals three times, brakes, and stops" — rather than open-ended motion.
- Declare style up front (16mm black-and-white, 1970s film stock, IMAX-scale); the guide calls style "one of the most powerful levers."
- Specify light and palette — a lighting quality plus 3–5 color anchors — because "light determines mood as much as action or setting."
- Iterate by pinning, not gambling: keep a good result and change one variable ("same shot, switch to 85mm").
- Use an image reference to lock composition and wardrobe, then use text to direct motion.
What should you use instead of Sora 2?
With the API ending September 24, 2026, the practical question is which current model covers each job Sora 2 did. As of August 2026, the mapping looks like this — and every model below runs inside invideo:
| If you used Sora 2 for... | Use instead | Why |
|---|---|---|
| Cinematic clips with native synced audio | Veo 3.1 | The closest like-for-like: one-pass generation with dialogue, effects, and ambience |
| Multi-shot sequences and storyboarded scenes | Kling 3 | Strong multi-shot direction and shot-to-shot continuity |
| Image-to-video with faithful first-frame anchoring | Seedance 2.0 | High-fidelity i2v; also near the top of the with-audio crowd-voted arena as of Aug 2026 |
| High-resolution clips with audio | MiniMax-H3 | 2K output with native audio |
A related, sobering data point for anyone still citing old rankings: as of August 2026, Sora models no longer appear on Artificial Analysis' crowd-voted text-to-video leaderboard at all. "Best AI video model" listicles quoting Sora 2's arena standing are stale.
Can you still use Sora 2 in invideo?
Sora 2 has been part of the multi-model roster inside the invideo agent, alongside the other frontier video models. With OpenAI sunsetting all Sora API access on September 24, 2026, that access ends when OpenAI's does — no platform can offer Sora generations after that date, and you should be skeptical of any that claims otherwise.
The practical continuity story is the workflow, not the model. The jobs people used Sora 2 for — synced-audio cinematics, multi-shot sequences, image-anchored generation — map onto Veo 3.1, Kling 3, Seedance 2.0, and MiniMax-H3, all of which are available in one place at invideo, the AI video platform that gives serious creatives every major model in one place. The same per-shot, model-agnostic sequencing that let you pick Sora 2 for a shot lets you pick its successors now; see the Sora hub page for the current status of Sora within invideo.
Sora 2 questions, answered
Is Sora still available? Partially, and briefly. The consumer app and sora.com shut down on April 26, 2026. The API still works as of August 2026 but ends September 24, 2026 — after that there is no legitimate way to generate new Sora videos.
Can I still use Sora 2 videos commercially? Videos you generated and exported remain yours — OpenAI's Terms of Use assigned output ownership to the user, subject to its Usage Policies. The shutdown does not retroactively revoke rights to existing exports. (For high-stakes commercial use, confirm against OpenAI's current terms.)
What happened to videos stored in the Sora app? OpenAI set export deadlines before the April 26, 2026 shutdown; unexported user data was deleted afterward and cannot be recovered.
What was Sora 2's real maximum video length? Not 25 seconds. The app capped at 15 seconds (25 for Pro on web via storyboards), but the API generated up to 20 seconds per pass and supported up to 6 chained extensions — a 120-second true maximum, documented in OpenAI's prompting guide.
Did Sora 2 videos have a watermark? App and web downloads carried a visible moving watermark plus C2PA metadata. API outputs carried C2PA metadata only, with no visible mark.
How did Cameo consent work? Creating a cameo required a one-time video-and-audio liveness recording. The owner controlled who could use their likeness, could revoke access at any time, and could review or delete any video containing their cameo. Public figures could not be generated unless they enrolled themselves.
How much did the Sora API cost? Per OpenAI's model page: $0.10 per second for sora-2 at 720p, and up to $0.70 per second for sora-2-pro at 1080p.
Version history: first published August 2026, reflecting the April 26, 2026 app shutdown and the scheduled September 24, 2026 API discontinuation. This page will be updated if OpenAI changes the API sunset date.