PAN 2 and Google DeepMind’s Genie 4 both pair a world model with language-model reasoning, but they offer different levels of access and control. A comparison by OrcaRouter argues that knowing which model does the reasoning matters as much as seeing the generated world.
Watch Desk analysis
What happened
OrcaRouter’s comparison describes PAN 2 as an open-weight system that generates a video clip from a starting frame and an action description. Genie 4, presented as a research preview, instead creates a live scene that can be steered with keyboard input. The OrcaRouter comparison says PAN 2 identifies its reasoning model and allows it to be changed, while Google’s description of Genie 4 names models involved but does not explain how to substitute or separately assess them.
The practical differences are stark: OrcaRouter says PAN 2 can be run as a controlled batch process, while Genie 4 is built for interactive exploration and includes generated audio. PAN 2’s weights are available under Apache-2.0, according to the comparison; Genie 4 access is limited to a research preview and waitlist, with no self-serve price stated.
Why it matters
A generated scene is only part of a world model’s job. If a language model is deciding what happens next, identifying that model makes it easier to understand the system and test how its reasoning affects the result. PAN 2’s disclosed, swappable reasoning component gives researchers more room to run controlled comparisons; Genie 4’s live interaction offers a different kind of experiment, but with less public detail about its reasoning setup.
These are not ready-made substitutes. OrcaRouter says PAN 2’s documented hardware requirements are substantial, while Genie 4’s access and performance details remain limited. One is easier to inspect and run under your own control; the other, for now, is the more immediately compelling place to explore.
Our read
The comparison’s useful insight is that “world model” covers more than a video generator. The reasoning model, the way a person steers the system and the ability to inspect or change its components all shape what it is good for. If you want to build repeatable experiments, PAN 2 looks like the more practical starting point. If you want to interact with a persistent scene, Genie 4 is the more intriguing prospect, assuming you can get access.
OrcaRouter’s account is a technical comparison, not a shared independent test of the two systems. Its architectural distinctions are useful; performance claims and practical fit still need testing against the task at hand.
What to watch
- Whether Genie 4 publishes more detail on its reasoning models, access and performance.
- Whether researchers test PAN 2 with different reasoning models on comparable tasks.
- How well either system maintains coherent scenes over longer interactions.
Discussion spark: For research on AI-generated worlds, would you prioritise an interactive system that is difficult to access, or an open model whose components you can inspect and change?
Sources and evidence
- PAN 2 vs Genie 4: Who Is the Language Model in the Loop? – OrcaRouter (11 November 2026, 15:17 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.