Decart's Oasis 3 and the wager on 'world models'

A world model is not a chatbot or an image generator. It is an AI that builds an interactive environment you can act inside — one that responds as you move through it, generating the next frame based on your actions rather than retrieving a stored clip. Decart's Oasis 3, launched 10 June 2026, is the most commercially pointed version yet: a model that spins up photorealistic driving scenes in real time from a text prompt, served via API. The ambition Decart is selling is not simply the demo — it is the platform.

Feed Oasis 3 a natural-language description and it generates a multi-camera driving scene — one front-facing view, two side — rendered photorealistically and explorable in real time. Weather, road surface, traffic density, lighting conditions: all variable, all promptable. The primary use case is autonomous-vehicle training, where teams need to rehearse edge-case scenarios at scale — black ice at 3 a.m., a cyclist appearing from behind a parked lorry, a flooded junction — that are rare enough never to be reliably encountered on real roads and too dangerous to stage repeatedly.

Decart prices the API at roughly $0.02 per second of simulated time. That is cheap enough that a team can generate thousands of scenario variants overnight; it is not free, which means Decart is trying to build a real business, not a demo. The company says robotics and broader 'physical AI' applications are the next expansion, once the driving domain is proven.

The strategic framing CEO Dean Leitersdorf is pushing is deliberately reminiscent of the early OpenAI API era: 'the first usable world model that people can actually program on top of'. The claim is not that Oasis 3 is the most impressive world model ever built — it is that it is the first accessible one, with pricing, an API, and a focused domain that makes it immediately actionable for a paying customer base.

That is a structurally significant distinction. Language models became a platform because developers could build products on them cheaply, with no ML expertise required. If world models follow the same trajectory — and Decart is making a $300m-backed bet that they will — the value accrues to whoever owns the layer everyone else's physical-AI applications run on. That makes the developer ecosystem more important than the demo reel, and the pricing more important than the benchmark score.

Multiple outlets that stress-tested Oasis 3 flagged real-world limitations in the simulations. This matters more here than in almost any other category. A photorealistic world model for entertainment can have visible glitches — that is fine. A world model being used to train a self-driving car that will subsequently operate on public roads cannot. The sim-to-reality gap is precisely the failure mode that AV training must guard against: a model that learns from slightly wrong physics, slightly wrong depth, or slightly wrong edge-case behaviour does not fail in the simulator — it fails on a real road at 60 mph.

'Photorealistic' is not 'physically accurate'. Decart has not claimed full accuracy; the caveats in the coverage are honest. What matters is whether the company can close the gap fast enough that the simulations teach correct behaviours rather than plausible-looking ones. The ~$300m raise at ~$4bn — with Toyota and Nvidia among the backers — buys runway to get there. Toyota's presence in the cap table is notable: an AV-adjacent manufacturer investing in a world-model startup is a signal of strategic intent, not just financial yield.

The comparison with language models is instructive here too. Early LLM APIs hallucinated constantly — that was tolerable for text generation and fatal for medical diagnoses. The world-model version of that same problem is a simulator with plausible-but-wrong physics. Which domains can tolerate imperfect simulation, and which cannot, will determine how quickly Decart's platform thesis can actually be tested at scale.

The table above maps what Oasis 3 is now against what the thesis requires. The honest read: the right product at the right time, if the fidelity improves.

Two things will determine whether the platform bet lands. First, developer adoption: whether AV and robotics teams actually integrate Oasis 3 into their pipelines, or treat it as a demo and stick with bespoke in-house simulators. Second, fidelity progress: whether Decart's next releases close the sim-to-reality gap enough that the simulations are safe to use for safety-critical training, not just exploration.

The robotics expansion is worth watching separately. Driving is a well-defined domain — camera angles, road physics, a finite set of actors. General robotics is vastly harder: unstructured environments, contact physics, manipulation. If Oasis 3's architecture generalises cleanly to manipulation tasks, the world-model platform thesis becomes much larger than AV training. If it does not, Decart has built a well-funded specialist tool in a category that is already served by the AV industry's mature in-house simulators. The answer will arrive in the next 12–18 months of developer usage, not in the launch press release.