Decart's new world model can simulate hours of photorealistic driving - with some caveats

A bet that 'world models' will get their own developer ecosystem

Decart unveiled Oasis 3, its latest interactive world model that spins up photorealistic, drivable environments in real time. The headline strategy is distribution: by shipping API access from day one, Decart wants to seed a developer community around world models the way OpenAI did with language models.

What it does and who it's for

• The first target is autonomous-vehicle companies that need to simulate rare driving scenarios at scale, with robotics and other "physical AI" uses planned next.
• Access is priced at $0.02 per second, with custom enterprise pricing, and it generates multi-camera (one front, two side) views for training and testing.
• Crucially, developers can generate scenarios infinitely rather than poke at limited demos - handy for AV teams hunting edge cases.
• It's built on Decart's Lucy video model and a community the company says already tops 100,000 developers.

There's serious money behind it: Decart recently raised $300 million at a near-$4 billion valuation, drawing strategic investors like Toyota, Adobe, and eBay alongside existing backer Nvidia - all potential customers.

The cheaper-to-run edge

Decart's claimed advantage is efficiency. Its DOS (Decart Optimization Stack) tunes models down to the hardware across Nvidia, Amazon, and Google chips, which the CEO says makes them more than an order of magnitude cheaper to run than rivals - so cheap the company claims it has burned through far less than $100 million in its life.

The caveats are real

TechCrunch's hands-on testing found the limits quickly. The model sets up a convincing opening scene, but coherence unravels as you move: a prompted New York street drifts into a generic Western cityscape, and turning around can replace the original intersection entirely. Controls are loose, and the model doesn't simulate physics - vehicles pass right through each other - which the CEO calls a major research problem tied to a lack of crash data. The root challenge is architectural: Oasis 3 is auto-regressive, generating one frame at a time (roughly 8,000 tokens each), so its context window fills almost instantly and memory is hard to hold. Decart is racing to extend that memory, and frames the current rough edges the way early LLM builders did - as a starting point for developers to build on.