OpenAI made waves earlier this year when it revealed Sora, a video-generating model that could allegedly take any snippet of text and turn it into a realistic video. The system has been in private testing since then, but OpenAI is now opening the velvet rope. The company's servers are being hammered, but as soon as that's worked out, you'll be able to take the video bot for a spin.
Sora operates on the same generative AI foundations as text-based large language models (LLMs) like ChatGPT, Claude, and Gemini. However, the input tokens are used to generate video pixels instead of text. The examples provided back in February, as well as the demos available today, are impressive and head-and-shoulders above what other commercially available video generators can do.
Due to high traffic, OpenAI has disabled account creation for Sora, which effectively keeps anyone from trying the system. However, OpenAI says Sora access will be included with existing ChatGPT accounts. The tool offers various options when creating videos, including length, resolution, and aspect ratio. This all boils down to the number of compute credits it costs to process the input, which is shown in the interface.
Many observers wondered if the delay in launching Sora was because of high compute costs, and it does look like Sora will be very spendy. The video above shows a job that costs 100 credits, and the number you get varies based on your subscription tier:
ChatGPT Plus subscribers ($20 per month, 1,000 credits) can create 50 videos per month at 480p "square." Subscribers can also choose to create a smaller number of 720p videos, which are at least 3x more expensive than 480p videos, but they're all limited to 5 seconds.
The new ChatGPT Pro account ($200 per month (10,000 credits) offers 1080p resolution and videos up to 20 seconds long. You get up to 500 of the smallest videos per month (480p square, 5 seconds long), but the HD videos can get very expensive at longer lengths. It would be possible to run through all 10,000 credits with just five 20-second 1080p videos (2,000 credits each).
All of these numbers are for "priority" videos, which are generated immediately. Pro subscribers can get an unlimited number of "relaxed" videos that you may have to wait to see.
OpenAI says the first public version of Sora is far from perfect. It often ends up creating unrealistic physics and awkward-looking complex actions. The company is also still working to make the model more affordable and faster. OpenAI also notes, somewhat distressingly, that it's counting on the community to help develop norms and safeguards to ensure Sora is used responsibly. The videos do at least have C2PA metadata embedded in them to show they are AI-generated. There's also a visible watermark on videos, but Pro users can remove that.