Decart launched Oasis 3 on Wednesday, an interactive world model that generates photorealistic driving environments in real time via API at $0.02 per second. The two-year-old startup is targeting autonomous vehicle companies first, then robotics and broader physical AI, but its larger bet is on developers — opening API access from day one in a deliberate echo of how OpenAI seeded the LLM ecosystem. Oasis 3 arrives a few weeks after Decart raised $300 million at a valuation of nearly $4 billion.
The funding round, which closed weeks before the Oasis 3 launch, brought in Toyota, Adobe, and eBay as strategic investors, with existing backer Nvidia participating. CEO Dean Leitersdorf said all three new investors are potential customers, with demand spiking in e-commerce, live streaming, and physical AI.
Decart already runs a developer community of more than 100,000 around Lucy, its real-time video foundation model. Oasis 3 is built on that same foundation and represents the company's push into physical-world simulation. Enterprise pricing varies by use case, on top of the published $0.02-per-second rate.
Key facts
- 01Decart launched Oasis 3, a real-time interactive world model generating photorealistic driving scenes via API at $0.02 per second.
- 02The two-year-old startup raised $300 million weeks ago at a nearly $4 billion valuation, with Toyota, Adobe, eBay, and Nvidia backing.
- 03Oasis 3 generates roughly 8,000 tokens per frame at tens of frames per second, pushing hundreds of thousands of tokens per second.
- 04Decart claims it has burned drastically less than $100M in its lifetime, and already counts over 100,000 developers using its Lucy video model.
- 05Rivals include Google's Genie 3, World Labs' Marble, and video-generation work from Luma and Runway.
The world model arena is crowded. Google released Genie 3 in research preview last year, Fei-Fei Li's World Labs shipped Marble for commercial use, and video-generation startups Luma and Runway have been retrofitting their physics-aware video models into interactive world models. Decart's pitch is that its vertical optimization stack — what it calls DOS, the Decart Optimization Stack — runs models far cheaper on Nvidia, Amazon, and Google hardware than the alternatives.
That cost edge is how Decart can let users wander a generated world for hours rather than offer capped demos. Leitersdorf says the company has burned drastically less than $100 million in its lifetime, an unusually low number for a frontier-model startup at a $4 billion valuation.
Oasis 3 produces multi-camera environments — one front-facing and two side-facing feeds — accurate enough for AV companies to train and test perception stacks against. The infinite-generation feature is aimed squarely at the edge-case problem: autonomous vehicle teams need volume in rare scenarios, and a model that quits after thirty seconds is not useful.
The trade-off shows up in coherence. In testing, the model produced a convincing New York City street from a single prompt, but the scene drifted as the camera moved — the city lost its specificity, intersections vanished, and the car drove through other vehicles. Leitersdorf attributes the physics gap to training data: there is drastically more footage of good driving than of accidents. He calls physics consistency a major research problem the team is cracking now.
The memory architecture is the bottleneck. Oasis 3 is auto-regressive, generating one frame at a time and looking backward to decide what comes next — compute-intensive by design. Decart is researching longer context windows to store millions more tokens and compression techniques to fit memory into fewer tokens.
Leitersdorf says the next version will let users seed worlds from a video of a real environment rather than a single image, which he expects to partially solve the consistency drift. World models as a field, he conceded, are still early — closer to GPT-2 than to GPT-4 in maturity.
The honest read is that Oasis 3 is the most photorealistic open API-accessible world model shipping today, with real flaws that the next version is explicitly designed to address. Object permanence and physics consistency are not trivial — they are the hard problems of the field — but every world-model vendor faces them, and Decart's pricing and uptime advantage gives it room to iterate in public. Hallucinated geometry in a driving sim is a bug to fix, not a verdict on the technology.
The strategic question is whether Decart's developer-first playbook actually replicates the OpenAI flywheel. Leitersdorf is betting that 100 developers building 100 unexpected applications in the next three months will define the category, the same way third-party builders defined what LLM APIs were for. If that bet works, the $4 billion valuation looks early rather than late — Decart becomes the default substrate for AV simulation, robotics training, and any application that needs a programmable physical world. If it doesn't, the company is a fast, cheap renderer competing with deeper-pocketed rivals on a problem the labs themselves want to own.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
