news

World Labs opens Atlas to everyone with two web apps and a public API, and still lists no price

The world model that claimed a 75 to 94 percent human preference over five rival video models in September is now self-serve, with those numbers still unreproduced and no published rate card.

By Daniel Ochoa · 13 November 2026

World Labs moved Atlas out of closed partner access on November 12, opening the world model to anyone who signs in at its own domain, with two browser apps and the API endpoints that power them. What did not ship alongside it is a price, a model card, or any outside reproduction of the preference numbers the company has been quoting since the model was announced on September 1.

Access had previously been limited to unnamed partners through a request form, and the model was reachable mainly through Marble Studio rather than having a front door of its own. Viewfinder, the image-based app, had been in closed beta with a waitlist.

Viewfinder works from up to 64 photos, Chisel works from blocked-out geometry

The two apps run on the same model but take different inputs. Viewfinder reconstructs a real environment from as many as 64 overlapping photographs, or generates a plausible surrounding scene from a single image, and exports spatially consistent frames, video clips or 3D Gaussian splats. Chisel is the geometry-first route: you block out a layout from a text prompt or reference images, move objects around, and every camera view updates with the structure before you render images, video or splats.

Two further tools are listed with the release, one converting splats into meshes for games and interactive environments and one for relighting splats. The developer API is available without a partner agreement and exposes endpoints including atlasGenerate, atlasChisel, images2PosedRGBD and splats2Mesh, which covers scene generation, video, posing and 3D reconstruction, and is pitched at agents and existing pipelines including Blender. A real-time variant sits alongside this as a research preview, streaming generated worlds from cloud GPUs to a browser at up to 15 frames per second while the user drives the camera.

The preference figures are World Labs' own, and two accounts give different ranges

The quality claims carried into this release come from the September announcement rather than from anything new. In blind human preference tests on camera-controlled generation, World Labs reported that raters picked Atlas 75 percent of the time against MiniMax H3, 81 percent against Gemini Omni Flash, 86 percent against Happy Horse 1.1, 93 percent against FLUX 3 and 94 percent against Seedance 2.5. Secondary accounts do not agree on the span: most report 75 to 94 percent, while at least one summary states the range as 81 to 93 percent. On sparse-view 3D reconstruction, the company reported a mean absolute-relative pointmap error of 25.3 against 28.7 for the next-best baseline, with the baselines reproduced in-house and closed commercial systems left out of the comparison.

Two caveats travel with those margins. Several comparisons used different test conditions for Atlas than for the competing methods, in a direction that favored Atlas, and Atlas takes a camera path as a native input while the video models it was measured against were given camera motion as text. That makes at least part of the gap a measurement of input format rather than of generation quality.

No paper, no model card, no rate card

World Labs has still not published an academic paper or a model card for Atlas, and no independent group has reproduced its reconstruction or preference results. Pricing for the newly public apps and API is not stated either, so the one number that would let a studio plan around Atlas, cost per generation, is the number nobody outside the company has.

Sources: World Labs, World Labs blog, VP Land.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.