OpenAI’s Sora 2 Brings Realistic Physics, Audio, and Continuity to AI Videos, Launches App

AI

by Nino LeitnerOctober 1^st, 2025

Subscribe to News - Software

OpenAI has unveiled Sora 2, a flagship video and audio generation model that aims for far better physical realism, multi-shot consistency, and fine-grained control, and it launches alongside a new social iOS app built around “cameos” of your own likeness.

OpenAI frames Sora 2 as a leap from early “it works” demos toward a system that better simulates cause and effect. The model is designed to model failure states and obey more of the everyday physics filmmakers expect, while adding synchronized dialogue and sound design that can follow prompts with greater precision.

What is new in the Sora 2 model

OpenAI says Sora 2 handles scenarios that previously broke video generators, such as Olympic-level gymnastics and backflips on a paddleboard that respect buoyancy and rigidity. Where older models might teleport a missed basketball into the hoop, Sora 2 lets it rim out or rebound from the backboard, an example of the system modeling outcomes rather than forcing success. The company positions this as a step toward a general-purpose world simulator. 

Controllability and multi-shot consistency

Prompts can now span multiple shots while maintaining scene state, character continuity, and blocking. OpenAI highlights improved instruction-following across realistic, cinematic, and anime styles. For production use, that means fewer continuity breaks when evolving a scene, and greater reliability when iterating on coverage.

Native audio: dialogue and sound effects

Sora 2 can generate background soundscapes, speech, and effects in sync with the visuals, all within a single promptable system. For quick previs, animatics, or social pieces, that reduces round-trips to separate audio tools.

Cameos: inject real-world likeness and voice

A major new capability lets users “upload” themselves. After a short capture, the model can insert a person’s appearance and voice into generated scenes with notable fidelity, and OpenAI says this generalizes to any human, animal, or object. Control remains with the cameo owner, who can revoke access and remove videos that include their likeness.

The Sora app: social media for AI videos

Alongside the model, OpenAI is launching an invite-based iOS app called “Sora”, only available for US and Canada at launch. The feed is tuned toward people you follow and creations likely to inspire your own videos. OpenAI claims they emphasize non-addictive design, optional personalization controls, wellbeing checks, and parental controls that can limit scroll and manage direct messages for teens. The company says monetization will be limited initially to paying for extra generations when compute is constrained.

Availability and access

The Sora iOS app rolls out first in the United States and Canada, with expansion planned. Access opens via in-app signup. After receiving an invite, users can also access Sora 2 through sora.com. Sora 2 will start free with generous limits that remain subject to compute availability. ChatGPT Pro users get access to an experimental, higher-quality “Sora 2 Pro” in ChatGPT, with support coming to the Sora app and an API release planned. The previous Sora 1 Turbo remains available, and past creations stay in users’ libraries.

Safety questions remain

OpenAI outlines consent controls for cameos, provenance measures, automated safety stacks, and scaled human moderation, with stricter defaults for teen accounts. The company has published additional safety and feed-design documents to detail guardrails and philosophy. 

How exactly they’ll make sure you are only using your own image for cameos remains to be seen, we’re guessing you need to prove your identity somehow with a photo ID. If not done properly, this could prove problematic with a surge in fake news videos disguising as real.

What Sora 2 could do for filmmakers

For previs and pitchvis, Sora 2’s attention to basic physics and continuity could cut the time needed to iterate on blocking, lensing, and stunt beats. The native audio layer makes quick tone pieces possible without separate tools. Cameos could help directors rough in performances with collaborators for timing and eye lines, then swap to actors as projects evolve.

For VFX and post, the realism claims are promising, yet the industry will still need to test how the model holds up on stitching, motion coherence in long takes (if you can even make them long), and edge cases such as complex occlusions, water interactions, and fine-grained hand articulation.

As with all generative video, licensing, credits, and provenance remain essential considerations on professional work, yet OpenAI usually trained theirs models on whatever online content they could get their hands on – so all of that remains questionable.

Will these tools meaningfully slot into your previs or concepting workflow this year, or does the state of generative video AI appal you? Let us know in the comments.