On September 22, 2026, PixVerse announced R2, the next step of its real-time world model after R1 in January 2026. The idea is simple to state: instead of typing a prompt and waiting for a fixed clip, you step into a world that keeps running, and your input changes it as it goes.
What the announcement actually claims
PixVerse describes R2 as a world that keeps running rather than a clip that finishes. Input persists: a prompt, an action or an audio cue updates the running state and shapes what comes next, and new input can arrive while the world is still generating. Text, references, audio and action controls all go into the same session. The Game Engine that PixVerse first launched in July 2026 now runs on R2.
What the post does not say is just as worth noting: no latency figure, no frame rate, no resolution, no session length, no price and no API details.
What PixVerse says changed
The company describes R2 as a scaling release. Its own wording is that the model can take on more data, tasks and input types while still running in real time. Two architecture terms appear: Omni Causal AR, a single causal model that is trained continuously, and a real-time acceleration layer that compresses that same model for live use instead of training a separate fast one.
That is a claim about method, and the post points to a separate technical report for the details. Nothing in the announcement itself lets a reader check it.
What the public examples show
The post shows three cases. In an explore mode, a player moves a character with WASD and the camera with the arrow keys, while prompts add elements or change the environment. Zero Mark, an interactive film game by creator Xiaolongbao, reacts differently when a creature is offered a dragon or a leaf. Eve, a character by creator Jade Wu, is said to keep identity and memory across a conversation.
These are curated demonstrations chosen by the company. They show the intended behavior. They do not say how often it holds up, or on what hardware.
Where R2 sits in the timeline
Readers who know the tool only as a text-to-video app will find this a different product. Our earlier PixVerse text-to-video tutorial covers the clip workflow from February 2024, which is not what R2 does.
What is still unknown
The post links a collection of R2 worlds at world.pixverse.video but does not describe sign-up, pricing or device requirements. Figures such as model size, benchmark scores and failure cases are not in the announcement either. Until the technical report or independent runs fill that in, treat R2 as a stated direction with a few staged demos, not a tool you can plan a project around.