It Was One of AI Video's Biggest Names — Now OpenAI Is Shutting It Down
Sora 2 Review 2026: OpenAI's video-and-audio model closed its consumer app in April and retires its API on September 24 — days from this review. Here's what it could actually do, and where it fell in the year since launch.
When Sora first appeared in February 2024, it looked like a genuine leap — a model that could turn a text prompt into a coherent, cinematic video clip. Sora 2, released September 30, 2025, pushed that further: synchronized dialogue and sound effects generated alongside the video itself, meaningfully better physical consistency, and a character system that let creators reuse a consistent figure across multiple generations. For about six months, it was one of the most talked-about names in AI video.
That window has closed. OpenAI shut down the Sora consumer web and app experience on April 26, 2026, and the Sora API — including both sora-2 and sora-2-pro — is officially deprecated and scheduled to retire entirely on September 24, 2026. This review is being published days before that shutdown takes effect. What follows is what Sora 2 was, what it could do, and where it stood by the time OpenAI decided to end it — not a recommendation to build anything new on it.
The Shutdown Timeline
OpenAI hasn't published a detailed financial postmortem explaining the decision. Independent reports have linked it to the high compute cost of video generation and a broader reallocation of resources toward other products, but those remain outside analyses rather than an official OpenAI explanation — worth knowing before treating any specific dollar figure circulating online as confirmed. Users had a window to export their Sora content from the sunset page before it closed; after that window, OpenAI states the data is permanently deleted.
What Sora 2 Could Actually Do
Synchronized audio generation
Sora 2 generated dialogue, sound effects, and environmental audio directly alongside the video rather than as a separate step — a genuine differentiator at launch, when most competing models produced silent clips.
Improved physical consistency
OpenAI's own example: a missed basketball shot could bounce realistically off the backboard rather than teleporting into the hoop or vanishing — a meaningfully better handle on how objects move and interact than the original Sora, though OpenAI itself acknowledged the model still made mistakes.
Characters and video extension
A short 2-4 second clip could become a reusable "Character" asset — person, animal, mascot, or object — carried across later generations, capped at two characters per video. Separately, a completed video could be extended up to 20 seconds per pass, up to six times, for a theoretical maximum around 120 seconds, though Characters and image references couldn't be combined with Extensions.
Targeted video editing
Rather than "inpainting" in the broader sense, the API supported targeted modification of an existing video through a specific prompt — OpenAI's own guidance noted that narrow, focused edits produced better results than trying to change many things in a single pass.
Real-person likeness was restricted by design: human Characters were disabled by default in the API, image inputs containing human faces were rejected, and generating real people — public figures included — wasn't permitted through the API. Sora also built in visible watermarking and C2PA provenance metadata, though OpenAI itself noted that metadata can be lost or stripped during common operations like format conversion, resizing, or screenshots — provenance signals, not a guarantee of detectability.
Pricing (Final Documented Rates)
| Model | Resolution | Rate | 20-second cost |
|---|---|---|---|
| Sora 2 | 720p | $0.10/second | $2.00 |
| Sora 2 Pro | 720p | $0.30/second | $6.00 |
| 1024p | $0.50/second | $10.00 | |
| 1080p | $0.70/second | $14.00 |
These were OpenAI's officially documented API rates — presented here as historical record, since the API retires September 24, 2026, days after this review's publication. No free API tier existed. It's also worth separating the API from the consumer product: ChatGPT Plus and Pro subscriptions once included Sora generation as part of the consumer app, but since that app shut down in April, those plans no longer offer a path to Sora access at all.
Where Sora 2 Stands Now
This is the part the story mostly left out of coverage at launch: how Sora 2 aged against a market that kept moving. Artificial Analysis currently lists Sora 2 (December snapshot) at 1097 Elo in text-to-video with audio, and Sora 2 Pro at 1073 — both measured against thousands of blind preference votes. For comparison, the current top of that same leaderboard sits at 1243 (Wan 3.0), 1238 (Gemini Omni Flash), 1225 (MiniMax H3), and 1220 (Seedance 2.0). Sora 2 isn't a bad model by that data — it's a model the field caught up to and passed within about a year, which is a different, more honest story than either "amazing" or "bad."
Strengths & Weaknesses
Sora 2 vs Wan 3.0 vs Current Leaders
| Model | Status | Text-to-video Elo (with audio) | Max length |
|---|---|---|---|
| Sora 2 | Deprecated, API retiring Sept 24 | 1097 | 20 seconds |
| Sora 2 Pro | Deprecated, API retiring Sept 24 | 1073 | 20 seconds |
| Wan 3.0 | Active | 1243 | 30 seconds |
| Gemini Omni Flash | Active | 1238 | Varies |
| Seedance 2.0 | Active | 1220 | Varies |
The comparison itself is the story here more than any single number. A model that was arguably category-defining a year ago is now both discontinued and measurably behind the current leaders on the same independent benchmark — a reminder of how fast this specific category has been moving through 2026, not a verdict on whether Sora 2's underlying engineering was good.
Is There Still a Reason to Use It?
Practically, no — not for new work. With the consumer app already closed and the API retiring within days of this review's publication, there's no path to start a workflow on Sora 2 today that wouldn't need to be rebuilt on another platform almost immediately.
Where it's still worth understanding: as a reference point for how the AI video category evolved — synchronized audio, physical consistency, and character continuity all became baseline expectations across the field partly because Sora 2 demonstrated they were possible, even after the product itself was retired.
Expert Editorial Opinion
The old version of this review said, in effect, "start with Plus, upgrade when the output earns its keep." That advice is not just outdated — it's advice for a product that no longer exists in any form a new user could act on. That's worth sitting with rather than quietly editing away: recommendations about fast-moving AI products have a shelf life, and this is as clean an example as exists of one expiring completely.
What's more interesting than the shutdown itself is the trajectory it caps off. Sora 2 launched into a field where synchronized audio was a genuine differentiator; by September 2026, it's closer to table stakes, and models like Wan 3.0 and Seedance have pushed past it on independent benchmarks while adding capabilities — 30-second single-pass generation, document-to-video input — that weren't part of the conversation a year ago. OpenAI's decision to retire Sora rather than keep iterating on it suggests the company concluded the video-generation race wasn't where it wanted to keep spending compute, at least for a standalone consumer product — a strategic read the company hasn't confirmed in detail, but one the timeline supports better than any single leaked cost figure does.
The clean-narrative version — "Sora was killed because it was too expensive" — is plausible but unconfirmed, and worth resisting as a headline claim. What's actually confirmed is simpler and still notable: a technically important, genuinely influential AI video model went from launch to full shutdown in under twelve months, in a category still young enough that a year is most of a product's entire public lifespan.
Final Verdict
This score reflects Sora 2's practical standing today, not a judgment of its engineering at launch. A model that was genuinely influential in late 2025 now carries a low score for one overwhelming reason: it's about to become entirely unusable, with no self-hosting option, no ongoing product, and no path forward for anyone evaluating it as of this review. Historical technical merit is real and acknowledged throughout this review — it simply isn't what this score is built to measure once a product has an announced shutdown date days away.
| Dimension | Weight | Score | Why |
|---|---|---|---|
| Technical quality | 30% | 6.5/10 | Genuinely strong for its 2025 launch window — synchronized audio, physical consistency — but now measurably behind current leaders on the same Artificial Analysis benchmark |
| Price-to-value | 25% | 2.0/10 | Irrelevant going forward — the API retires days after this review, so no ongoing value can be delivered at any price |
| Maturity & documentation | 20% | 2.5/10 | Was well documented while active; now officially deprecated with a fixed end date, which dominates this dimension regardless of past documentation quality |
| Ceiling & flexibility | 15% | 3.0/10 | Had real capability (extension, characters, editing) but restricted access to human likeness, and the discontinued status removes any forward flexibility entirely |
| Honesty of positioning | 10% | 7.0/10 | OpenAI announced the shutdown clearly and gave an export window, even without publishing a detailed cost rationale |
Weighted total: (6.5×0.30) + (2.0×0.25) + (2.5×0.20) + (3.0×0.15) + (7.0×0.10) = 4.10/10 — Score band: below 6.0, "Do not build on this — not a quality judgment, a status one: the product is being retired."
❓ Frequently Asked Questions
🔗 Related ToolRadar Reviews
Official shutdown notice and documentation: OpenAI Help Center. If you have content stored in Sora, verify the current export window status directly, as OpenAI's stated deletion timeline is enforced after that window closes.
Comments
Post a Comment