OpenAI released Sora 2, a new version of its text-to-video generation model, on 30 September 2025 [source]. The announcement marks the second major iteration of the system first previewed in February 2024.

According to the provider's blog post, Sora 2 generates video from text prompts with improved temporal consistency and motion quality compared to the original release. OpenAI states the model can now produce clips up to 20 seconds in length at 1080p resolution, an increase from the previous 5-second limit at lower resolution.

The update introduces what OpenAI describes as "enhanced understanding of physics and object permanence," addressing criticisms of the first version's tendency to generate physically implausible motion. The provider also claims reduced latency, with generation times dropping from several minutes to under 60 seconds for most prompts.

Sora 2 remains available only to ChatGPT Plus and Pro subscribers. OpenAI has not disclosed the model's training data sources, architecture changes, or compute requirements. The announcement does not address outstanding questions about content provenance or watermarking, which were raised during Sora 1's limited release.

The provider states that safety mitigations include prompt filtering and output screening for policy violations, though specific technical details were not provided. OpenAI notes that generated videos include C2PA metadata for provenance tracking, a feature not present in the original version.

No independent benchmarks or third-party evaluations were available at the time of release. The announcement follows similar video generation launches from Google, Runway, and Stability AI in recent months.