Sora 2: Examples, Critiques, and Early Impressions

Besides that, the model promises a handful of other improvements. Let’s walk through each claimed improvement and look critically at a few examples.

More physically accurate

Sora 2 is supposed to better understand concepts like weight, balance, object permanence, and cause-and-effect. If someone misses a basketball shot, the ball is meant to bounce off the rim—rather than teleport magically into the hoop like we’ve seen in older models. Let’s analyze this example:

The physical dynamics look pretty convincing for most of the video, but if you look at the very last frames (the last 0.5 seconds), you’ll see that the skater’s legs stretch into an impossible and almost deforming shape.

When the cat drops from the skater’s head, the fall feels weightless and unnatural. Then, at the very end, the cat does an impossible pirouette.

Another major problem with this video is that we have audio continuity while missing video continuity. Listening to the announcer speaking over the microphone in the arena, we can deduce that the action and time are continuous because there’s no cut in the sound, but the 10-second video is actually composed of three different shots that clearly happened at different times. It’s clear to me that Sora 2 doesn’t have much of a clue about basic montage principles.

Capable of simulating failure

Another subtle (but very useful) upgrade is that Sora 2 can now generate realistic mistakes. Older models mostly rendered successful actions—someone always lands the jump, always scores the goal. Sora 2 tries to simulate failure, which is harder than it sounds. For AI, “messing up” in a realistic way often requires a deeper internal model of physics and human behavior.

Let’s check an example:

I actually quite like this video, it almost convinces me that the fall is real. The only thing that breaks the illusion is the impossible balance toward the end. It’s debatable, but I think the man should have fallen much earlier.

That said, the ability to generate mistakes has a lot of potential. The most obvious use case that comes to mind is when you need to film a dangerous shot. Instead of using an actual person, maybe you could try AI to avoid complications and cut down on budget as well.

Better continuity

OpenAI says Sora 2 can obey complex prompts across multiple shots while maintaining the same character, lighting, and general world state. So if you ask for a man in a red jacket to enter a building, walk down a hallway, and pick up a newspaper from a desk, the jacket should stay red, the hallway should stay consistent, and there shouldn’t suddenly be two newspapers or three arms.

This is the most relevant example I’ve found for this feature:

This example packs an impressive six shots into a single 10-second video. What really sells the continuity is the sound—the background chatter, fireworks, and dialogue are all impressively well generated. I can’t judge the semantics because I don’t understand the language being spoken, but acoustically, it works.

However, once you take a closer look at the visuals, things start to get confusing. The distance between the characters shifts too much—sometimes they’re far apart, then suddenly they’re touching hands just seconds later. The camera also jumps the axis in what feels like a random way, making it really hard to understand the geography of the space and the position of the characters.

That said, the face consistency is perfect, so I’ll give the model credit for that. The lighting is also consistent, and the black-and-white color grading holds up across all the shots. 

Flexible in style

OpenAI claims the model can shift fluidly between realistic, cinematic, and animated styles—including anime—while preserving motion and identity. This is useful for creators who want stylized output without completely losing control of the scene.

I have to say, this is perhaps the example I like the most from everything they’ve shown. There are both things that work and things that don’t work, but story-wise, this is a beautifully crafted scene, in my opinion, and this is mostly because it has a functional emotional dynamics.

The world around the protagonists is a happy one: people are enjoying some kind of night festival, fireworks light up the sky, and there’s a sense of celebration. But in the middle of all that, the two main characters seem to be having a very difficult conversation—it feels like they’re on the verge of an irreversible breakup.

I don’t understand the language, so I can’t say for sure what they’re saying, but the emotional tone comes through clearly. The music is very well chosen—slightly upbeat, in line with the setting—and the close-up at the end, with the fireworks reflected in the female character’s eyes, is what really drives the point home.

If I had to complain about something, it’s that the people around them don’t move at all. But this could also be an intentional choice—we’re seeing a subjective view of the scene, where the world around them has stopped for a moment because they’re so caught up in their own tense world that, subjectively, the outside world has ceased to exist for them.

Sora: The Social App

Alongside the model, OpenAI also launched a new iOS app called Sora, which adds a full social layer to the video generation experience. Right now, the app is the main way to get access to Sora 2—and so far, it’s invite-only and limited to users in the U.S. and Canada.

I guess OpenAI wants to see what happens when people start remixing each other’s content, inserting themselves into scenes, and treating AI video more like a social language. Whether this becomes the next TikTok is a different question (I think the chances are zero), but the mechanics are interesting.

How does the Sora social app work

At its core, the Sora app is a feed-based platform built for remixable AI video. You scroll through short clips generated by other users, many of which include real people’s likenesses through a system called cameos (more on that below). If you want to create your own, you can either write a new prompt from scratch or remix an existing video using someone’s cameo—with their permission.

You can “like,” remix, repost, or follow users just like you would on any other social platform. The feed itself is personalized, but OpenAI says they’re not optimizing for watch time. Instead, they claim it’s tuned to maximize creation and inspiration, and they’ve added well-being controls, including daily generation limits for teens and parental overrides via ChatGPT.

The app also includes a natural language–controlled recommendation system, meaning you can steer the kind of content you see in plain English (e.g. “show me funny videos with animals” or “less surreal stuff”).

Cameo features

This is probably the most novel feature of the app. Cameos let you record a short one-time video and audio clip of yourself, which the model then uses to generate a full-body likeness—voice, appearance, expressions, and all. Once your cameo is verified and uploaded, you can insert yourself into any scene or allow friends to remix you into theirs.

OpenAI’s framing here is that you remain in control of your likeness at all times:

• You decide who can use your cameo.
• You can revoke access or remove any video that includes you.
• You can view all videos (including drafts) where your likeness appears.

It’s a clever system, but it also raises a lot of questions—about consent, copyright, moderation, and how this scales. OpenAI says they’re building out human review teams to catch edge cases like bullying, and that stricter permissions are in place for younger users.

Why the Sudden Rush on AI-Generated Media?

Just like a recent Stanford report pointed out, AI hasn’t really disrupted that many industries yet—with one clear exception: media.

Source: State of AI in Business 2025 Report

I don’t think that’s just a statistical blip. I’ve spoken directly with advertising producers, and they’re taking this shift seriously. Many are actively experimenting with AI to reduce production costs, speed up timelines, and avoid the logistical headaches that come with filming on set.

You can see the same trend playing out publicly. Scroll through LinkedIn on any given day, and you’ll find creators posting AI-generated short films, ads, and concept trailers. A good example is this Kalshi ad—an actual ad for a real company—entirely produced using AI:

That’s why I think companies like OpenAI, Google, and Meta are suddenly going full throttle on AI-generated media. If there’s one domain where generative AI is already proving useful—and where people are actively spending time and money—it’s here. So it makes sense that labs are racing to build platforms, ecosystems, and social layers around video.

Just last week, Meta launched Vibes, a new short-form feed of AI-generated videos, which seems to be going after the same space as OpenAI’s Sora app. The pattern is pretty clear: capitalize on media and entertainment now, while the technology still isn’t good enough to reliably disrupt other, more regulated industries like healthcare or finance.

In short, video is the lowest-hanging fruit—and everyone’s trying to grab it first.

Conclusion

While I want to stay critical, it’s worth acknowledging how far these AI models have come. Just remember that in early 2023, the best we could get was the Will Smith eating spaghetti video. Now we’re talking about native audio, emotional continuity, and full-blown social remix apps where you can drop yourself into a sword fight on a pirate ship or a surreal cooking show hosted by your dog.

That said, Sora 2 still struggles with spatial coherence, basic editing logic, and subtle physical realism. And OpenAI’s decision to gate access behind an invite-only iOS app doesn’t exactly help its reputation with pro users looking for serious tools. There’s potential here, but there’s also friction.

Still, it’s clear that AI-generated media is the space to watch.