Decart launched Oasis 3 this week, a real-time interactive world model that generates photorealistic driving scenes from a single text prompt, available via API in what TechCrunch reported as an exclusive. It produces a multi-camera view built for autonomous-vehicle work, one front-facing and two side-facing streams, generating roughly 8,000 tokens per frame auto-regressively at tens of frames per second. It runs across NVIDIA, Amazon, and Google hardware through Decart's own optimization layer, priced at two cents per second of generated world. CEO Dean Leitersdorf says it costs more than an order of magnitude less to run than anything else in the field, and that it is the first world model usable enough that people can actually program on top of it.
The caveats are not buried, they are the point, and the headline says so. Over a long run the world drifts: New York streets melt into generic city, and if you turn around and come back, the place you left is gone and replaced. Directional control slips out from under you. And the model does not simulate collisions, cars drive straight through one another, because, Leitersdorf says, the training data holds far more good driving than crashes. He calls the collision gap a major research problem the company is cracking now. The honest summary is that long sessions feel dream-like and disjointed, photorealistic frame by frame but not coherent or physical across time.
That gap matters precisely because of the pitch. The use case is autonomous-vehicle training, and the single most valuable thing an AV simulator can produce is the dangerous moment: the near-miss, the cut-off, the collision you want the car to learn to avoid. Oasis 3 renders the street beautifully and cannot yet render the crash physically. So the bet is photorealism-first, physics-later, cheap and programmable now. That is a different bet from the rest of the field: DeepMind's Genie 3 is a general world model gated behind a $200-a-month tier, Waymo's world model is trained on 50 million real autonomous miles, NVIDIA's Cosmos targets physical reasoning directly. Decart is two years old, raised a $300 million round at a roughly $4 billion valuation with Toyota, Adobe, eBay, and NVIDIA in, says it has burned well under $100 million total, and already runs Lucy, a real-time video model, for a 100,000-plus developer community.
For the thread we have been tracking, this is the first commercial, and first honestly-caveated, data point in world-models-as-training-substrate, the lane Genie 3 opened two days ago. The substrate is now shippable, priced, and an order of magnitude cheaper than expected, which is the part that will pull builders in. But the caveat is the real signal: usable and physically faithful are not yet the same thing, and for embodied training that distinction is the whole game. A simulator where cars phase through each other can teach a perception stack what a street looks like, but not teach a planning stack what to do when another car does not yield. Decart's roadmap, video-based generation instead of image prompts, longer context, and memory compression for consistency, is aimed straight at the gap. Whether photorealism-first or physics-first wins for training embodied agents is now a live question with a price tag attached, and the first real answers will come from whoever tries to train on Oasis 3 and reports back.