Sora 2 Test Review: Where the Physics Breaks

I typed a prompt for a figure skater doing a triple axel, hit generate in my head, and then remembered I couldn't. That is the strange thing about writing a Sora 2 test review in the middle of 2026: the physics finally got good enough to argue about, right as OpenAI closed the front door.

Here is the honest version up front. Sora 2's headline promise is physics accuracy, and it half keeps that promise: it now models failure instead of faking success, but complex human motion, hands, and text still break in ways you can spot in seconds. I could not run my own generation for this piece, because Sora 2 sits behind a paid, logged-in account and the whole consumer app shut down on 26 April 2026. So this is a verified failure-map built from OpenAI's own claim and first-hand creator tests, not footage I cooked up.

What Sora 2 actually promised

Sora 2's big claim is that it models physical failure, not just success. In OpenAI's own announcement (30 September 2025), the team said older models were "overoptimistic" and would let a missed basketball "spontaneously teleport to the hoop," while Sora 2 makes it "rebound off the backboard," because "you must be able to model failure, not just success."

That single sentence is the whole pitch. A model that lets a shot miss naturally is closer to real physics than one that quietly cheats the ball into the net. On that narrow point, Sora 2 delivers, and it is a real step past Sora 1.

The trouble starts the moment you ask for a body doing something hard.

What I tested, and what I couldn't

I did not run a single Sora 2 generation for this review, and I want that on the record before anything else. Sora 2 is only reachable through a paid, logged-in OpenAI account, and the consumer app (both sora.chatgpt.com and iOS) was discontinued on 26 April 2026, per OpenAI's own Help Center. The Sora 2 API is still live for now, but it sunsets on 24 September 2026.

So instead of faking a test, I did what the honest version of this looks like: I pulled the physics claim straight from OpenAI, then checked it against documented first-hand tests from reviewers who ran real clips before the shutdown. Every fail below is sourced, not something I generated.

Where Sora 2's physics actually breaks

Sora 2 breaks most on limbs, hands, and text, which is exactly where physics accuracy matters to a human eye. The rebounding basketball works. The bodies around it do not always cooperate.

In DataCamp's hands-on test (1 October 2025), "the skater's legs stretch into an impossible and almost deforming shape." A falling cat looks "weightless and unnatural" and finishes with "an impossible pirouette." Two characters stand far apart, and "suddenly they're touching hands seconds later," while a background crowd stays frozen in place. I have watched a lot of AI skaters grow a third leg, and it still gets me every time.

Hands are their own category of pain. Asked to count from one to five on-screen, the model held up five fingers while saying "3." Object interaction was clean only about 60% of the time even at Pro quality, according to unite.ai's review. And here's the weird part: text warps just as badly, so a prompted "DHL" uniform came out as "a series of illegible lines rather than a clear logo."

The root cause is the same one that makes AI image tools stumble over fingers, which is worth understanding before you trust any of them: these models predict the next pixel from patterns, not from a real physics engine. It is the same guesswork we ran into when we dug into what Midjourney actually is. Fast Company documented the original Sora rendering gymnasts who "land on three legs or an extra head" for the same reason.

What Sora 2 genuinely does well

Sora 2 is strong wherever the scene is stable rather than acrobatic. Native synchronized audio is the standout: dialogue, ambience, and sound effects come out of the same generation, which almost no rival did at launch.

It also holds faces and characters steady across shots, keeps lighting and color consistent, and carries an emotional thread through a short scene. And the physics honesty is real progress: a missed shot bounces off the rim instead of magically dropping in. See for yourself in any before-and-after against Sora 1, and the jump is obvious.

I would put its physics roughly at the level of Google's Veo 3. Good enough to fool a casual scroll. Not good enough for anything with a spinning body in frame.

The door just closed, so trying it is hard now

The most useful fact about Sora 2 today is that a normal person can no longer just try it. OpenAI shut the consumer app on 26 April 2026 and set the API to sunset on 24 September 2026, reportedly because the app cost around $1 million a day to run against thin revenue, plus copyright and deepfake pressure.

Before the shutdown, access ran through ChatGPT Plus at $20 per month, with Sora 2 Pro sitting behind ChatGPT Pro at $200 per month. None of those doors open to Sora 2 anymore. Every clip it made also carried OpenAI's logo plus C2PA provenance metadata, a hidden tag that marks a file as AI-generated, so nothing left the tool anonymous.

If you want creative tools you can actually open today instead of a video model that is halfway out the door, our roundup of free Canva alternatives is a far more practical starting point.

Would I trust it for real work?

I would trust Sora 2 for mood, faces, and dialogue, and not for anything athletic. For a talking character, a calm product shot, or a short scene with steady motion, it is genuinely good and the audio saves you a whole editing step.

For sport, dance, precise hands, or readable text, it is a gamble you will lose more often than not. And since the easy way in is gone, that decision may be moot for most of us until the next model lands. My honest read: the promise was real, the execution is partial, and the timing turned a hot tool into a case study.

Here is a stress test you can run on any video model you do have access to, to find its breaking point in minutes:

A single figure skater performs a slow triple axel spin in an empty rink, arms tucked, then extends both arms and counts from one to five on their fingers to the camera. Their jacket has a clear readable logo. Even indoor lighting, one continuous 15-second shot, no cuts.

Watch the legs during the spin, count the fingers, and read the logo. That one clip tells you more than any spec sheet.