Leading Proprietary Models
last edited
by early this year, the leading proprietary video models formed a two-horse race. OpenAI’s Sora 2 and Gen-4.5 co-led on cinematic quality and realistic physics. The open ecosystem is covered at Open-Source Video Generation Landscape, while the parent hub is AI Video Generation — Organized Reference.
Sora 2 (OpenAI)
Sora 2 launched in September 2025 as a social iOS app. The model’s diffusion transformer uses “world modeling,” including object interactions in 3D space. The signature strength is physics: gravity, momentum, collisions, fluids, and light refraction convincing enough for elite gymnastics or a paddleboard backflip.
The native audio is synchronized, though generated as a secondary pass. Clips run 5–25 seconds. Pricing is $20 per month for Plus or $200 per month for Pro.
A $1B licensing deal was announced in December 2025. The model’s opt-out-by-default copyright stance, easily stripped watermarks, and “SlopTok” feed drew sustained criticism.
Gen-4.5
Gen-4.5 generates one-minute, multi-shot videos with character consistency, native dialogue, and background audio.
Reported tests show 30–40% fewer artifacts than competitors. Scaling to 4K is clean. A three-second clip takes about 18 seconds to process. Pricing is $12 monthly for the standard plan and $28 monthly for Pro.
Reported customers include major film studios, advertising agencies, gaming companies, and architecture firms. Adoption rose 300% year over year.
Its partners include a film studio, a cloud provider, a creative-software company, and NVIDIA. The image counterpart is Frontier Image Models.
Sources
Further details are available from OpenAI’s Sora 2 announcement and the Gen-4.5 announcement. Company research and release materials provide the original descriptions of model capabilities, access, and product positioning.