Decart’s new world model can simulate hours of photorealistic driving — with some caveats
Decart unveiled Oasis 3, a real-time photorealistic driving simulator exposed through an API. It targets AV testing, but long runs still break down.
Intelligence analysis by GPT-5.4 Mini
Oasis 3 is Decart’s bet that world models can become a programmable platform for developers, starting with autonomous driving simulation. It looks impressive and runs for hours, but the experience still loses coherence, physics, and object awareness over time.
Decart made a computer-made road world that can keep changing while a car moves through it, like a video game that invents the next scene on the fly. It looks very real at first, but after a while the world forgets details and starts to wobble.
Analysis
What Decart launched
Decart introduced Oasis 3, an interactive world model that can generate photorealistic driving environments in real time and expose them through an API. The company says it is first targeting autonomous vehicle developers who need to simulate rare driving scenarios at scale, with robotics and other physical AI use cases next.
Decart is also framing the product as a developer platform, not just a model demo. CEO Dean Leitersdorf told TechCrunch the goal is to make it the first usable world model people can actually build on top of, echoing how language models developed broader ecosystems. The startup says it already has more than 100,000 developers using its real-time video model Lucy, which helped form the foundation for Oasis 3.
The pitch
The company says Oasis 3 is priced at $0.02 per second, with enterprise pricing varying by use case. Leitersdorf also says Decart’s software stack makes its models more than an order of magnitude cheaper to run than competitors because the company optimizes across Nvidia, Amazon, and Google hardware.
That efficiency claim matters because the model is meant to be used repeatedly, not just shown in a short demo. Decart says the system can generate scenarios infinitely, which could help AV teams explore many edge cases.
The caveats
TechCrunch’s testing found the model can produce a strong opening scene, but the simulation degrades as the drive continues. A New York City prompt started convincingly, then drifted into a generic urban setting; when the car turned back, the original intersection was gone. The controls were also described as hard to manage, and the car sometimes drove through other cars, showing that physics still is not reliable.
Leitersdorf called the collision problem a major research issue and said the team is working on better memory, noting that each generated frame uses roughly 8,000 tokens. That helps explain the tension here: Oasis 3 is visually striking and potentially useful, but it still behaves more like a dream than a stable simulation.
Key points
- Decart launched Oasis 3, a real-time photorealistic world model for driving simulation, and made it available through an API.
- The company is initially targeting autonomous vehicle companies, then plans to expand into robotics and other physical AI applications.
- Decart says the model is cheaper to run than rivals because of its vertically integrated optimization stack.
- In testing, the simulation degraded over time, with scenes changing, controls becoming less responsive, and collisions not behaving realistically.
- Decart recently raised $300 million at a valuation of nearly $4 billion, with strategic investors including Toyota, Adobe, eBay, and Nvidia.
If Decart improves memory and physics, Oasis 3 could become a useful place to test driving edge cases that are hard or unsafe to recreate in the real world. API access from day one could also help build a real developer ecosystem around world models.
The model may stay impressive to look at while still falling short as a serious simulator if it keeps losing scene continuity and object awareness. If the physics problems remain, developers may use it mostly for demos rather than training or testing.
