Decart's world model is real — and narrow on purpose

Launched 10 June 2026: a world model that generates photorealistic, multi-camera driving scenes from plain-language prompts, in real time, via API at roughly $0.02 a second. Target buyer: AV teams who need to rehearse rare and dangerous scenarios at scale without waiting for them to happen on real roads. CEO Dean Leitersdorf calls it the 'first usable world model that people can actually program on top of' — which is a platform pitch, not just a product pitch. He wants developers to build on top of Oasis 3 the way the OpenAI ecosystem grew on top of GPT-3.

Decart's earlier Oasis demos were open-ended 'explore any world' showcases — impressive, general, and with no clear paying customer. Oasis 3 makes a deliberate choice: drop the breadth, buy the depth. Driving is a constrained domain — defined camera geometries, road physics, a finite set of actors — which means a world model can achieve useful fidelity faster there than in unstructured environments.

This also gives Decart a paying customer base from day one. Autonomous-vehicle companies already budget heavily for simulation — both mature in-house simulators and a range of specialist vendors compete for that spend. The question is not whether AV teams will pay for high-quality simulation — they will, and do — but whether Oasis 3 offers something different enough to win a share of that budget: namely, the ability to generate any scenario from a text prompt rather than authoring it manually.

The comparison with the early LLM API era is apt, and Leitersdorf is right to reach for it. Before the OpenAI API, language-model capabilities existed in research — nobody was building products on them at scale, because access was constrained and pricing was unclear. The API flipped that: suddenly thousands of developers could build products, pricing was legible, and the ecosystem compounded. If world models follow the same arc — and there is no guarantee they will — the company that owns the API layer during the early adoption window ends up with an enormous structural advantage.

The narrowing is not a failure of ambition. It is the thing that makes the ambition viable.

Outlets that tried Oasis 3 flagged real-world flaws in the simulations. Normally 'demo has rough edges' is a forgivable early-product note. Here it is the entire risk category. If you are training a self-driving car against a simulation, the gaps between sim and reality are precisely where the car learns the wrong lesson. A world model that is 95% photorealistic can teach a 5% wrong behaviour, and that error surfaces on a real road at motorway speed, not in the test environment.

This is not a reason to dismiss Oasis 3 — it is the thing to track. The relevant question is not 'does it look impressive' but 'is the sim-to-reality transfer accurate enough that the behaviours it trains are the behaviours you want in the physical world'. That answer is not yet clearly yes.

So file it correctly. Oasis 3 is genuinely interesting, appropriately scoped, and backed by people with real strategic skin in the game. The bet that pays off is not the pretty driving clip in the launch video — it is whether developers build enough on top that Decart becomes the layer everyone else's physical-AI stacks run on. That bet depends on fidelity improving fast enough that the simulations are safe for safety-critical training, not just exploration.

The platform ambition is real. The fidelity problem is real. Both can be true, and the next 18 months of developer usage will settle which one dominates the story.