OpenAI Announces Sora 2 — Its Next-Generation AI Video and Audio Model
OpenAI announced this week the release of Sora 2, its new video and audio generation model capable of creating highly detailed and realistic videos from user-provided text prompts. According to reports, the model outperforms Google’s Veo 3 in both realism and overall performance.
For now, Sora 2 is available only in the United States and Canada, and by invitation only. OpenAI stated that it plans to expand access to other countries soon.
This new version follows its predecessor, Sora 1, bringing significant improvements.
Key Enhancements in Sora 2
Generates 30- to 60-second video clips with sharper details than Sora 1.
Improves temporal consistency, reducing flickering and frame instability.
Produces more physically accurate motion, with better interaction between gravity and objects.
Fully integrated into the ChatGPT Pro environment, offering more accessibility than Google’s Veo 3.
However, when it comes to native audio (such as character dialogue and environmental sounds), Google Veo 3 still leads, since Sora 2 does not yet generate audio.
Opinion: AI Video Generation Still in Its Baby Steps
In my opinion, this AI video generation technology is still in its infancy — we’re several years away from seeing truly perfect, cinematic-quality videos.
As of 2025, users must still spend hundreds of credits (and quite a bit of money) generating multiple clips, many of which fail or come out with visual defects.
Creating even a short film remains a labor-intensive process.
For instance, a 40-minute short film equals 2,400 seconds. If you were to use 5-second clips, that would require 480 individual video clips.
That’s hundreds of credits and dollars spent on an AI generator — and it could take weeks to complete.
If each 5-second clip costs, on average, 60 credits, the total would reach 28,800 credits.
It’s probably best to wait another year or two until a truly capable AI tool emerges — one that can create longer, more coherent videos.
For example, Google’s Veo 3 can already generate clips up to 60 seconds long, but the longer the video, the more inconsistencies tend to appear — objects disappearing, people with three hands, or physics that simply don’t make sense.
Silvio Guerrinha
0 Comments