A peak is not a frame rate you can plan around
Start with the wording, because the wording is the specification. "Up to 15 fps" describes the best case the company is willing to print. It does not describe the rate on a complex scene, the rate ten minutes into a session, or the rate on whatever GPU a given visitor's session happens to land on. No average, no minimum and no test setup accompany the number.
The missing figures matter more than the present one. Resolution decides whether 15 frames a second looks like a world or like a preview of one, and the announcement does not give it. Readers who know Atlas's headline specification may supply 1440p from memory. The September specification belongs to Atlas's offline, camera-path-controlled video generation: up to one minute at 1440p. World Labs has not said that Frameboy's live stream operates at that resolution. Transferring it to a live stream the company optimized for speed is exactly the move the announcement invites and nowhere supports.
Latency is the third absence. For a camera you steer, the delay between moving and seeing decides whether the loop feels direct, and the post does not state it.
What the demo actually lets you do
The public half is narrower than the headline. Frameboy offers scenes World Labs chose: some built from a single image, with Atlas inventing the viewpoints around it, and some navigable reconstructions of real places from multiview captures. You can explore those well. You cannot bring a photograph of your own street and walk through it in the public demo.
That second job belongs to Viewfinder, where a scene is built frame by frame from your inputs, each new frame carrying color and depth, and the Frameboy overlay is then switched on to explore it live. Viewfinder is in closed beta. So the workflow most readers will picture when they hear "explore your own world in real time" is, for now, a waitlist.
None of this makes the preview minor. Putting generation inside the navigation loop changes what a world model is for: scouting a viewpoint for a film shot, checking a game level, or testing where a robot can go are all tasks where the value sits in trying the next angle immediately. World Labs makes that case itself, and it is a good one. It is a case about iteration speed, which is why the unreported numbers are not footnotes. They are the product question.
The honest way to try it
The demo is public, so the check is cheap. Pick one of the curated scenes, turn sharply rather than drifting, and watch two things: whether the frame rate holds through the turn, and whether fine detail survives movement or arrives a moment after you stop. Those two observations will tell you more about where real-time world models stand than the peak figure does.
Then hold the result against the cost sentence the company wrote about its own system. World Labs says Frameboy currently requires a powerful cloud GPU and is not cheap compared with local rendering methods. That helps explain the infrastructure burden of the preview, but the company does not disclose per-session costs or say that GPU expense is the reason access remains limited to curated scenes. Until someone publishes the resolution, the latency and the session cost together, "real time" here names a direction the work is traveling, at a speed nobody outside World Labs has yet measured.
Sources
- [1] World Labs, "Frameboy: Atlas in Real-time" (November 5, 2026; read first-hand at Gate 1: Frameboy "streams frames from cloud GPUs to your browser at up to 15 fps as you control the camera" and serves a variant of Atlas "highly optimized to achieve the fastest, most efficient inference possible"; Frameboy offers a curated list of scenes seeded from a single image or multiview captures of real places; Viewfinder "is currently available to test in closed beta; you can sign up to our waitlist," with each new Atlas frame including color and depth; "Frameboy currently must run on a powerful GPU server in the cloud and is not cheap compared to local rendering methods"; Atlas "takes concepts from RTFM and scales them up" and is "a much larger model trained on orders of magnitude more data"; the post states no streaming resolution, latency figure or price)Read source
- [2] World Labs, "Atlas: A World Model for Spatial Intelligence" (September 1, 2026; read first-hand at Gate 1: camera-controlled generation "outputting up to 1 minute of video at 1440p" — Atlas's offline specification, not a Frameboy streaming figure; Atlas "will power future versions of Marble and other products")Read source
- [3] World Labs, Research & Insights blog index (read first-hand at Gate 1: Frameboy dated November 5, 2026; Atlas dated September 1, 2026; RTFM dated October 16, 2025)Read source
- [4] Unite.AI, "World Labs Releases Frameboy for Real-Time Atlas Exploration" (November 5, 2026; curated scene list including Stanford Quad, Sutro Baths and Palace of Fine Arts; Viewfinder color-and-depth workflow)Read source