No—Decart’s Oasis did not make game engines obsolete. Its 2024 demo showed that a generative model could turn keyboard input into game-like video at a reported 20 frames per second. It did not demonstrate the persistent world state, consistent rules, or precise control expected of a general-purpose game engine. Decart’s current Oasis 3 is aimed instead at physical-AI simulation, including autonomous-vehicle scenarios.
What the original Oasis was
Decart and Etched introduced Oasis on October 31, 2024 as an interactive, Minecraft-like video experience. Players could use keyboard input to move, jump, pick up objects, and break blocks. The project described apparent physics and rules, but explicitly said it had no conventional physics engine: “There is no physics engine; just a foundation model.” The release included code, weights for a locally runnable 500-million-parameter model, and a larger hosted demo. Those were research and demonstration artifacts, not a consumer game-engine product. Decart and Etched’s launch documentation describes the project and its architecture.
Technically, Oasis combines a spatial autoencoder with a latent diffusion backbone, both Transformer-based. It generates video frames autoregressively: the model uses recent frames and user input to predict what comes next. That creates an interactive experience without the familiar arrangement of a simulation updating explicit values for, say, a block’s position, an inventory slot, or a collision. The launch authors also noted that errors can compound in autoregressive generation; long-context generation and dynamic noising were intended to reduce temporal instability.
What Oasis actually proved
Generative video could respond at interactive speed
The clearest result was throughput. Decart and Etched reported that the original Oasis generated 20 frames per second, or a new frame every 0.04 seconds. They contrasted that with their estimate that contemporary text-to-video systems could take 10–20 seconds to generate one second of video. That was the project authors’ comparison, not an independent, matched benchmark across equivalent systems. Still, the demo made a meaningful point: a model could produce a stream of game-like images in response to player input quickly enough to feel interactive.
#1 Best Overall
Input-conditioned images can imitate the consequences of actions
Press a key, provide the model with recent visual context and input, and it generates another frame. That can make movement and interaction look game-like. But generating a plausible next image is not the same thing as executing a simulation whose state and rules are explicitly represented and updated. The distinction matters when an action must have a precise, repeatable outcome or when an object must remain where it was left.
Decart CEO Dean Leitersdorf made a similar contrast in a Sequoia Capital interview, describing conventional engines as worlds where “The world is very consistent. You can really make things very accurate.” He argued that generative systems could make worlds easier to modify. That points to a real trade-off, not proof that one approach has already replaced the other.
Rank #2
What it did not prove
It did not demonstrate a persistent, dependable game world
Oasis generated successive frames; the demo did not establish that it maintained Minecraft-like world state over time. Its own launch documentation identified fuzzy distant video, temporal consistency for uncertain objects, domain generalization, inventory precision, object control, and long-context memory as areas for further development. In a hands-on report, WIRED observed the environment changing when the viewer looked away. Interactive frame generation therefore should not be confused with a persistent and reliable simulation.
“No physics engine” did not mean better or more reliable physics
The phrase describes the demo’s architecture: learned visual prediction stood in for explicit physics and game rules. It does not mean Oasis supplied a more dependable physics engine. Where an application needs exact collisions, stable object locations, deterministic behavior, or repeatable results, the demo did not show that generated video could meet those requirements.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsIt did not show that developers can stop using engines
Game engines also provide tools and workflows for representing state, enforcing rules, authoring scenes, debugging behavior, and controlling how a world changes. Oasis did not show that those jobs had disappeared. A plausible role for generative models is to complement an engine or help create more varied environments; which responsibilities belong to each remains an open design question, not a result established by the demo.
How Oasis 3 changes the story
As of October 4, 2026, Decart positions Oasis 3 as an API-accessible world model for physical-AI simulation, initially emphasizing autonomous vehicles and also naming robotics applications. Its official Oasis 3 page describes prompt-defined settings, action-conditioned feedback, synchronized multi-camera views, and possible applications involving drones, off-road vehicles, maritime settings, and humanoid manipulation.
Decart says Oasis 3 responds in under 200 milliseconds and generates at 22 FPS at 512×768×3. These are current vendor performance claims, not independent benchmark results. The figures and product description indicate a new application focus; they do not establish that the 2024 game-like demo became a replacement for conventional game engines.
TechCrunch reported on June 10, 2026 that Oasis 3 was available by API for $0.02 per second, with enterprise pricing dependent on use case. Pricing and access can change. In the reporter’s hands-on experience, a prompted driving scene initially matched the request but lost thematic and geographic continuity as it moved, and the controls could be difficult to direct. That account is a useful reminder that even a model aimed at physical-AI scenarios has consistency and controllability questions.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Best Value
How to compare a world model with a game engine
They should be judged on the jobs an application needs, rather than on whether generated images look convincing in a short clip. The available sources do not provide a controlled head-to-head evaluation, but they point to the relevant questions:
- State: Is the world represented explicitly and persistently, or inferred from recent visual context?
- Consistency: Do objects and locations remain stable across time, movement, and changes in camera view?
- Control: Can developers and users produce precise, repeatable outcomes?
- Authoring and variation: How much work is needed to build a world, and how easily can its appearance or setting change?
- Latency and cost: What response time and compute cost does the application require?
Oasis made real-time generative interaction more tangible. It did not settle how these systems should divide work with conventional engines, or whether a world model can satisfy the reliability and authoring needs of a general-purpose game.
What the result means for game development
The useful takeaway is neither that Oasis was merely a visual trick nor that engines are finished. It demonstrated a different path to interactive, game-like output, with the possibility of easier variation. Its documented weaknesses show why a convincing sequence of frames is not yet equivalent to a world developers can reliably author, inspect, and control. Oasis 3’s shift toward physical-AI simulation makes world models an adjacent development direction, not evidence that conventional game development has ended.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
