Summary: Decart released Oasis 3 on June 10, 2026, an interactive world model exposed through an API and built for robot control and driving scenes. The company cites 22 FPS, three synchronized cameras, sub-200-millisecond latency, and $0.02 per second pricing. TechCrunch’s launch-day testing still found drifting scenes, unresponsive steering, and vehicles passing through each other, so physical consistency remains unfinished.
Decart — a startup that got its start in 2023 — put out Oasis 3 on June 10, 2026: a world model you interact with and reach through an API. Hand it a text prompt and it renders a driving scene that looks like camera footage, updating in real time and continuing to evolve in a closed loop as the robot inside it acts. Decart’s own tagline calls it “the first world model accessible by API, live from day one” — a framing CEO Dean Leitersdorf pushed further at launch: “Every robot will be trained inside a world model. That era starts today.”
The takeaway: world models are leaving the demo stage and becoming compute you pay for by usage. Anyone building autonomous vehicles finally gets a published rate card for a machine that manufactures long-tail situations — snowy nights, tunnels, blocked cameras.
Inside Oasis 3
Underneath Oasis 3 are two pieces of Decart’s own infrastructure: Lucy, its real-time video foundation model, and DOS (Decart Optimization Stack), its inference stack. The design is an autoregressive scheme called “Live Stream Diffusion” — each frame is produced one after another, with previous frames and the robot’s actions serving as the conditioning inputs. The original Oasis listened to a keyboard and mouse; Oasis 3 is built around robot control instead, taking action signals directly, and its geometry-aware design emits three synchronized camera streams (one facing forward, two on the sides) — the same layout a real AV perception stack would see.
A text prompt steers the environment, and generation can run in a closed loop with no ceiling; per the company, drift gets corrected even in extended sessions. NVIDIA’s Physical AI Open Dataset on Hugging Face supplied the training data, and the model operates on CoreWeave clusters built from NVIDIA HGX B200 systems. Four intended applications get named: running online reinforcement learning inside the generated world, checking policies before they ship, harvesting expert training data through high-fidelity teleoperation, and recreating rare events that would be unsafe or impractical to set up in reality.
Specs, Speed, and Price Tag
The official numbers: 22 FPS, 512×768×3 resolution across three cameras, latency held under 200 milliseconds, with DOS credited for roughly a 100x jump in efficiency. The API costs $0.02 per second, with enterprise terms by use case. To make the scale tangible, Leitersdorf breaks it down: a single frame works out to about 8,000 tokens, which at tens of frames per second means the pipeline moves hundreds of thousands of tokens every second. A session with an environment can run for hours, and per Decart its community has already passed 100,000 developers, the bulk of them working on top of Lucy.
Where Hands-On Testing Fell Short
TechCrunch tried it on launch day and came away far less impressed. Prompted with New York City, the scene slowly morphed into an ordinary Western city, and the intersection it began with eventually disappeared altogether. The steering would stop responding for stretches at a time. Vehicles clipped straight through one another — real physics never enters the picture. The reviewer summed the experience up as a “dream-like, disjointed stream of consciousness” that slides into nonsense as time passes. Because of the autoregressive design, the context window fills up quickly; the team admits that longer memory and token compression remain open research, and a future release might let people kick off from a video instead of a single image.
On why the physics falls short, Leitersdorf is blunt: records of good driving vastly outnumber records of accidents, and that imbalance is, quoting him, “a major research problem that we’re cracking now.”
The Play Behind the Product
What the company is really after is an ecosystem. According to Leitersdorf, this feels like LLMs did in their early days, “when OpenAI invented the API for models,” and he expects “there’s going to be an entire developer community that emerges on top of this.” The money is in place too: the $300 million round unveiled in mid-May and led by Radical Ventures pulled in NVIDIA, Toyota Ventures, Adobe Ventures, and eBay Ventures, with Sequoia and Benchmark coming back for more; per TechCrunch the valuation sits near $4 billion and total funding has cleared $450 million. In Leitersdorf’s telling, lifetime spending by the company has stayed “drastically less” than $100 million.
Decart’s Spot in the World-Model Race
Competition across the 2026 world-model field is thick: Google’s Genie 3, World Labs’ Marble, plus Runway and Luma have all staked their ground — a lane we mapped in our opening-of-2026 outlook. Decart stands apart through vertical integration: it builds its own models and its own inference stack, one that runs on NVIDIA, Google, and Amazon silicon alike, and Leitersdorf says that setup leaves it “more than an order of magnitude cheaper” than competitors. Whether turning world models into APIs will follow the same commercialization arc LLMs did should become clear within a few quarters. For now, read “programmable world” as a heading rather than an arrival point: physical consistency is still the unfinished assignment every team on this track shares.