Open access, beta terms

Start with what kind of opening this is. It is product access, not an open release. Atlas is available as a hosted service, not downloadable weights. Customers remain subject to its service terms and cannot independently inspect or self-host the model weights.

The terms are visibly unfinished. The API reference lives on a beta host. The billing documentation states the credit conversion and a handful of Studio prices, then adds a sentence worth keeping attached to every number in this piece: the task rate card can change, and developers should read the current card in the platform rather than embedding prices in client code. The submission quotas and concurrent caps that decide whether Atlas fits a production pipeline are, on the rate card as reported by Radiance Fields, still to be announced, with higher throughput available by arrangement.

The published rates support provisional cost estimates for single jobs: a draft splat run at $0.50, a refined one at $5, a minute inside Frameboy at $0.50. What they do not settle is throughput. Unpublished quotas and concurrency caps leave production capacity uncertain, and the rates themselves can change. Budget a pilot from these figures. Do not budget a pipeline.

The benchmark to read carefully

World Labs' case for Atlas rests partly on human preference: in the company's camera-trajectory trials, third-party raters judging which model better followed the intended camera path preferred Atlas in 75 percent of trials against MiniMax H3 and 94 percent against Seedance 2.5, with the advantage growing as trajectories grew more complex. The company's own post describes the design plainly. Atlas received each camera path in its native camera format. The rival models, which do not accept camera geometry as an input, received the same moves described in text, and the post concedes that more sophisticated prompting might improve their results.

That does not make the result empty. Accepting a camera path as geometry rather than as a sentence is the product's central design choice. But the trial compares camera-path adherence under those different input methods; it does not isolate the effect of native camera conditioning, it is not an overall ranking of video quality, and no independent reproduction appears in the sources reviewed. The reconstruction results in the same post were run under a common protocol the company says it reproduced for all baselines; those baselines are specialist open-source models, and the evaluation remains the company's.

The underlying capability claims are narrower than the launch language around them. The one-minute, 1440p figure describes offline generation along a hand-designed camera path from a small set of reference images. Reconstruction, in the company's own account, typically needs only two or three images to stay faithful, can use more than a hundred, and invents plausible content wherever the inputs did not observe the scene. That invention is the feature when you are building a set. It is a liability when the output is treated as a record of a real place.

How to test it this week

The opening is cheap to probe honestly. Register, take one place you know well, and check the current platform rate for each job before running it: a Viewfinder reconstruction from your own photos, one camera-path render, and one splat-to-mesh export. Compare the reconstruction against the photographs at the edges, where the model had to guess. That comparison, plus the elapsed time and the credits each job consumed, tells you more about Atlas in an afternoon than the preference percentages do.

Then decide what the beta terms allow. Previz, virtual sets, design reviews and explorable drafts are tasks where plausible infill is acceptable and a changed rate card is an annoyance. For uses that require faithful real-world geometry, such as surveying, as-built documentation or robot training data tied to a specific place, plausible infill requires explicit validation, and an unpublished quota is a planning risk. Atlas now lets anyone run that test. It has not yet published the terms that would let a team skip it.

Sources

  1. [1] World Labs, "Atlas" product pageRead source
  2. [2] World Labs, "Atlas: A World Model for Spatial Intelligence" (September 1, 2026)Read source
  3. [3] World Labs, Atlas Developer API reference (beta docs)Read source
  4. [4] World Labs, Atlas billing documentation (beta docs)Read source
  5. [5] Radiance Fields, "World Labs Atlas Opens to All With a Splat API and Blender" (November 13, 2026)Read source
  6. [6] VP Land, "World Labs' Atlas World Model Opens to Everyone" (November 13, 2026)Read source