OpenAI Sora Turbo: A Very Expensive Slot Machine
OpenAI Christmas event: Day 3 of 12
Sora, OpenAI’s AI video model, was first announced in February 2024, though its release was initially withheld. Yesterday, on day 3 of their 12-day Christmas event, OpenAI finally launched Sora under a new version, Sora Turbo, along with its system card and a dedicated website.
If you remember Sora (and the two articles I wrote at the time), you already know a lot of what I’m sharing today. If you don’t—or if you want to know about the new stuff—then let me start by getting rid of that annoying adjective: The “Turbo” in the name means this version of Sora is faster—a crucial feature for a customer-facing product—compared to the version demoed in February. It doesn't mean this model is better (it may actually be worse as we’ll see) or that this is Sora 2 or something else.
So, Sora Turbo is just Sora adapted to face OpenAI’s massive 300-million-user base. But to achieve an increased generation speed, OpenAI may have engaged in unwanted trade-offs that are pissing off alpha users and early testers who tried the previous version. Some say the company has rushed an incomplete release.
Before getting into that, let me answer the most important question: Who can use Sora, and under what terms?
Plus users ($20/month): 50 videos, 480p or 720p (total of 1000 credits so you get fewer videos on higher resolution) and up to 5 seconds per video.
Pro users ($200/month): 500 videos, up to 1080p, up to 20 seconds per video. Also, you can download the videos without a watermark and can generate unlimited “relaxed” videos (which means low-priority generations).
People in the EU, Switzerland, and the UK haven’t been granted access (for now; perhaps never depending on the specifics of the regulation).
Demand has surged so quickly that OpenAI’s infra team hasn’t been able to keep up. They’re currently turning on and off sing-ups as they scale inference compute to accommodate paid users.
The Gist
I. What you need to know about Sora Turbo: Sora Turbo combines diffusion models (DALL-E 3) with transformers (ChatGPT), generating videos from text, images, or other videos. It offers features like remix, re-cut, loop, blend, and storyboard (OpenAI demoed all of them in February) but its outputs frequently lack physical consistency—doors behave unrealistically, objects morph, transitions falter, and there’s weak prompt adherence, making it feel like “a very expensive slot machine.”
II. Sora’s fundamental weakness: physics: OpenAI markets Sora as a “world simulator,” but the physical inconsistency of its outputs contradicts this claim. Critics like Gary Marcus argue that Sora’s issues stem from deeper flaws in reconstructing reality, casting doubt on whether scaling alone can bridge this gap. OpenAI believes this is just the beginning and trusting scale is what has gotten them so far so they’ll keep doing that.
III. Put away your sword of judgment: Judging Sora’s flaws harshly only from the quality of its outputs misses the point—this is nascent technology. GPT-1 was unremarkable compared to GPT-4. Likewise, the first iteration of AI video models sets a baseline, not a ceiling. Once Sora 2 is out, we could see the pattern and whether it looks good or nah. Better remain curious than judgmental at this time.
REMINDER: The Christmas Special offer—20% off for life—runs from Dec 1st to Jan 1st. Lock in your annual subscription now for just $40/year (or the price of a cup of coffee a month). Starting Jan 1st, The Algorithmic Bridge will move to $10/month or $100/year (existing paid subs retain their current rates). If you’ve been thinking about upgrading, now’s the time.