Sora 2: Amazing Model. Dubious Rollout.
OpenAI chose to release its top-tier video model inside a gimmicky social app.
TL;DR
Sora 2 is a genuinely impressive video model, but it’s baked into a social media app built for ultra-short meme clips, which undersells its creative potential.
What is it?
When OpenAI first teased its original Sora model back in February 2024, people went nuts. But Sora wouldn’t launch to the public until December, and by that time, it was well behind many top-tier video models.1
With Sora 2, OpenAI is convincingly back in the game:
There’s a lot to like about Sora 2.
It can handle complex physics, including intricate scenes like gymnastics and fight sequences. It can create videos in many different styles. And, like Veo 3, Sora 2 natively generates its own audio effects and dialogue.
While there are no third-party benchmarks or leaderboard rankings yet, Sora 2 feels roughly on par with Veo 3 (depending on the use case).
Here are three quick comparisons of my own:
Prompt #1: Over-excited blonde influencer is holding a smartphone with the Substack feed on it. She near-screams: “You guys have to check out Why Try AI. It’s so, so good. But what do I know? I don’t even exist!”
Veo 3:
Sora 2:
I don’t know why both models decided to throw in a semi-psychotic giggle at the end, entirely unprompted, but here we are.
Prompt #2: A drunken 1800s pirate tries to use a modern laptop but can’t figure it out. In frustration, he bangs the keys and says “Blast this shiny chest o’ letters—won’t open no matter how I pound it!”
Veo 3:
Sora 2:
I love the bonus details, from Veo 3’s exploding laptop to Sora 2’s off-script method acting improv.
Prompt #3: A horse wearing a top hat tap dances to music on its hind legs
Veo 3:
Sora 2:
Veo 3 looks more realistic, but I like how Sora 2 came up with an entire jingle. No true dancing on hind legs from either model, however.
The bottom line is that both are impressive and not too far apart.