The short version
● Same bet Both are world models - systems that predict what happens next.
● Praxis-1 Open weight, 27B, drives a real robot arm.
● Genie 4 Closed API, 720p at 24fps with native audio, for interactive worlds.
● The split One wants to control a machine. The other wants you to play.
Two world models landed eight days apart from companies with no overlap in their customer base, and the contrast says more about where this technology is going than either one does alone.
Runway built Praxis-1 to move a robot. Google built Genie 4 to generate a world you can walk around in. Both start from the same premise: a model that has watched enough video has learned how the physical world behaves.
Side by side
| Runway Praxis-1 | Google Genie 4 |
|---|
| Published | 11 November 2026 | 19 November 2026 |
| Purpose | Robot control | Interactive generated worlds |
| Weights | Open, per Runway | Closed |
| Access | Not stated | Interactions API, Gemini API |
| Size | 27B video foundation model | Not published |
| Output | Actions, via an action expert | 720p video at 24fps |
| Audio | Not applicable | Native, synchronised |
| Control | Joint commands | Keyboard plus live text prompts |
| Price | Your own hardware | Not published |
| Licence | Not stated | Standard API terms |
| Watermarking | Not applicable | SynthID on every output |
What each one proves
Praxis-1's claim is that you can skip robot training data. The action expert reads the video model's internal state rather than its frames, which means the physics knowledge was already there from pretraining and only the translation into joint commands had to be learned. Runway says this puts it top of RoboLab-120 among open-source robot foundation models.
Genie 4's claim is consistency over time. Generating one convincing frame is solved; generating the thousandth that still remembers what you changed in the fiftieth is not. Google says Gemini's reasoning holds the world state together while you alter weather, lighting and objects mid-session.
Both are missing the number that decides it
Neither company published the figure a builder actually needs. Runway claims leaderboard positions without a score, and states no download location or licence for a model it calls open weight. Google points at pricing documentation without publishing a rate, for a model generating 24 frames a second with audio.
In both cases that is the difference between a research result and something you can plan around. Price a realistic session before building on Genie 4; confirm the licence before building on Praxis-1.
Which one is relevant to you
They barely compete. If you have a robot, Praxis-1 is the only one of the two that does anything for you. If you are building a game, a simulation or an interactive demo, Genie 4 is, and the SynthID watermarking is worth knowing about before you composite the output into anything else.
The reason to read them together is the direction of travel. Video pretraining is turning into a general-purpose physics prior, and two companies with completely different businesses arrived at that within a fortnight.
FAQ
What is a world model?
A model that predicts what happens next in an environment. Both Praxis-1 and Genie 4 learn that from video, then use it either to control a robot or to generate an interactive scene.
Is Praxis-1 or Genie 4 open?
Runway describes Praxis-1 as open weight but publishes no download location or licence. Genie 4 is closed, available through the Gemini API.
How much does Genie 4 cost?
Google has not published a price. The announcement links to pricing documentation without stating figures.
Can Praxis-1 generate video?
It is built from a video foundation model, but what Runway published is robot control. The action expert reads internal state rather than generating frames to watch.
Which should I use?
They serve different jobs. Praxis-1 for robot control, Genie 4 for interactive generated worlds. There is almost no overlap.