How Sora 2 advances text-to-video-and-audio

OpenAI released Sora 2, a text-to-video-and-audio model designed to prioritize physical plausibility, multi-shot controllability, and synchronized dialogue and sound effects. Rather than treating each clip as an isolated best-effort synthesis, Sora 2 aims for simulation-grade output: better world modeling, consistent state across shots, and native time-aligned audio generation.

Improved realism and control

Sora 2 claims materially better world modeling, addressing common failure modes such as object “teleportation” by producing physically plausible interactions like rebounds on missed shots. It preserves state across multiple shots to enable instruction-following edits that remain coherent through a sequence, and it produces native audio that is time-aligned with generated video, including speech, ambient noise, and effects.

Cameos and the invite-only iOS app

OpenAI is launching a new invite-only Sora iOS app, initially in the U.S. and Canada, centered on a cameos system. Users record a short in-app video and audio to verify their identity and capture a likeness. Cameo owners control who can use their likeness, and they can revoke or delete any video — including drafts — that includes them. The app emphasizes social creation and remixing while keeping user consent central to any insertion of a verified likeness into generated scenes.

Safety, provenance, and launch restrictions

OpenAI outlines an iterative rollout with launch-time restrictions and provenance controls. At launch, the platform restricts image uploads that feature a photorealistic person and disallows all video uploads. Sora 2 does not support video-to-video at launch, blocks text-to-video generation of public figures, and prevents generations that include real people unless the cameo opt-in has been used. Additional classifier thresholds are applied when a real person appears.

All outputs include C2PA metadata for provenance and carry a visible moving watermark on downloads. OpenAI also uses internal detection tools for origin assessment, making provenance and traceability a core part of the rollout.

Parental controls and content philosophy

OpenAI introduced parental controls integrated via ChatGPT alongside Sora. Parents can opt teens into a non-personalized feed, manage direct message permissions, and control continuous scroll behavior. These controls align with the Sora feed philosophy of prioritizing creation over passive consumption.

Access, pricing, and roadmap

The Sora iOS app is available to download now and grants access by invite. Sora 2 is initially free under compute-constrained caps. ChatGPT Pro subscribers gain access to an experimental Sora 2 Pro tier on sora.com with an app rollout forthcoming. API access is planned after the consumer rollout, and existing Sora 1 Turbo content remains available in user libraries.

What this signals

Sora 2 represents a shift toward governed, production-oriented media tooling rather than raw capability demos. By combining consented likenesses, provenance metadata, and visible watermarks with conservative launch restrictions, OpenAI is prioritizing safety and traceability as it scales text-to-video-and-audio generation.