Sora

What is Sora?

Sora is OpenAI's text-to-video generation model, capable of creating realistic video from text descriptions. First previewed in February 2024 and publicly released in December 2024, Sora represents what some call the "GPT-1 moment" for video—the point where AI video generation started to actually work.

Capabilities

Video Generation:

• Up to 20 seconds of video
• Up to 1080p resolution
• Widescreen, vertical, or square aspect ratios
• Dynamic camera motion with consistent 3D space

Technical Features:

• Object permanence (objects persist across frames)
• Physics simulation (though imperfect)
• Video extension (forward or backward in time)
• Can simulate interactive environments like Minecraft

Evolution

Original Sora (February 2024)

• Preview release showing impressive demos
• Up to one minute of video at lower quality
• No audio
• Limited public access

Sora Turbo (December 2024)

• Significantly faster generation
• Public release to ChatGPT Pro and Plus users
• Plus users: 50 videos/month at 480p or fewer at 720p

Sora 2 (September 2025)

Major upgrade introducing:

Audio: Native synchronized sound generation—the biggest addition.

Better Physics: "If a basketball player misses a shot, it will rebound off the backboard." Mistakes now appear to be mistakes of the implicit agent being modeled, not physics violations.

Controllability: Follows intricate multi-shot instructions while maintaining world state consistency.

Style Range: Excels at realistic, cinematic, and anime styles.

Cameos: Users can include their own likeness in generated videos via a short verification recording. Accurate appearance and voice rendering.

"World Simulators"

OpenAI describes video generation models as "world simulators"—systems that learn implicit models of physics and object behavior. This connects to the broader research into world models that Demis Hassabis and others argue is necessary for AGI.

However, current video generators don't truly understand physics—they approximate it from patterns in training data. Generated videos may look realistic without being physically accurate.

Safety Measures

OpenAI has implemented safeguards:

• Watermarks: Visible markers on generated content
• C2PA metadata: Machine-readable provenance information
• Upload restrictions: Limited photorealistic uploads of people, especially minors
• Content policies: Preventing nonconsensual likeness generation

Access and Pricing

• ChatGPT Pro ($200/month): Highest quality and longest videos
• ChatGPT Plus ($20/month): 50 videos/month at lower resolutions
• iOS and Android apps: Released alongside Sora 2

Competitive Landscape

Sora competes with:

• Runway Gen-3: Professional video generation
• Pika: Consumer-focused video AI
• Google Veo: Google's video generation efforts
• Kling: Kuaishou's video model

Related Reading

• World Models - The concept Sora embodies
• Sam Altman - OpenAI CEO overseeing Sora development