OpenAI launches Sora Turbo, but physics remain wonky

OpenAI just pulled Sora out of research preview and slapped it onto sora.com as a proper product. It’s not the same model we saw back in February. They’re calling the new version Sora Turbo, and yes, it’s supposedly a lot faster. Plus and Pro subscribers get access starting now, though if you’re in the UK, Switzerland, or the European Economic Area, you’re out of luck for the moment. The pricing structure is pretty straightforward: Plus users can crank out up to 50 videos a month at 480p, or fewer if they want the bump to 720p. Pro users get ten times that, along with higher resolutions and longer durations. OpenAI is also teasing tailored pricing for different user types early next year, so they clearly see this as a tiered product, not just a single plan fits all.

The feature set is actually more interesting than the base model. Users get 1080p resolution, up to 20 seconds of video, and can switch between widescreen, vertical, and square aspect ratios. The storyboard tool is a standout — it lets you specify inputs for each frame, giving a level of control that text-to-video typically lacks. You can also extend, remix, and blend your own uploaded assets, or just generate entirely new content from a text prompt. A community feed shows off creations from other users, which feels like a direct nod to how Midjourney built its early momentum.

But let’s not get carried away. OpenAI is refreshingly blunt about the limitations, admitting the model “often generates unrealistic physics and struggles with complex actions over long durations.” That’s a significant caveat. Sora Turbo is faster, but it’s still tripping over basic movement and object permanence. They’re rolling it out anyway, arguing that society needs time to “explore its possibilities and co-develop norms and safeguards.” It’s a classic OpenAI move: ship early, gather data, iterate in public. The safety measures are about what you’d expect — C2PA metadata for provenance, visible watermarks by default, and an internal search tool to verify if content originated from Sora. Uploads of people are heavily restricted at launch, limited to a small pilot group while they refine deepfake mitigations.

The most aggressive guardrails are reserved for the ugliest abuse vectors. Child sexual abuse materials and sexual deepfakes are blocked outright, with reports sent to the National Center for Missing & Exploited Children. Red-teamers specializing in disinformation and illegal content were brought in before launch. Still, the whole thing feels like a controlled burn. OpenAI knows the physics are janky and the moderation isn’t airtight, but they’re betting that getting the tool into people’s hands now — with restrictions — is better than waiting for a perfect version that might never arrive. Whether creative communities embrace a tool this hobbled, or just get frustrated by the guardrails, is the real question.