World model APIs in 2026: which world models you can actually call from code

“World model API” is a search people run expecting a menu. There is no menu. Of the six world models that dominate the coverage in 2026, three can be called from code today, three cannot, and one video model that often gets lumped in is being switched off this month. This page, from an independent site not affiliated with any of the vendors, lays out which is which, what each API takes in and gives back, and what it costs. Every figure links to its source and carries a date — prices and availability move quickly in this category, and each vendor’s own documentation is the authority.

The short answer

Model	Vendor	Public API?	Output	Price basis
Marble 1.1	World Labs	Yes, World API since Jan 21, 2026	Gaussian splats (.spz, .ply), GLB mesh	1,500 credits per world (about $1.20)
Odyssey-2 Pro	Odyssey	Yes, developer.odyssey.ml	Real-time 720p video stream	Not covered here
MiniMax H3	MiniMax	Yes, MiniMax API, fal, ComfyUI; open weights	4 to 15 s video at 2K with audio	Per clip, varies by host
Sora 2	OpenAI	Closing Sep 24, 2026	15 to 25 s video	$0.10 to $0.50 per second
Atlas	World Labs	No, partner early access	1440p video, point clouds, splats	Not disclosed
Genie 3	Google DeepMind	No, Project Genie for AI Ultra (US)	Real-time 720p playable video	Subscription only
GWM Worlds 2	Runway	No, research preview	Real-time 720p video with audio	Not disclosed

Sources: World Labs, Odyssey, fal, OpenAI Help, Implicator, Wikipedia, Runway. All as of September 7, 2026.

If you need a 3D asset back, there is exactly one option on that list. If you need real-time interaction, there is one. If you need cinematic clips with sound, there is one that is open-weights and one that is closing. The list is the seven models that dominate 2026 coverage, not every vendor with an endpoint, but it is the set most teams are actually choosing between.

Three kinds of “world model”, three kinds of API

The phrase covers products that do different jobs, and the API shape follows the job.

3D world generators return geometry. You send a prompt, an image or a video, wait a few minutes, and download a scene you can open in a viewer or import into an engine. The call is asynchronous: create a job, poll it, fetch the export. Marble is the only public example. Atlas will be the second when World Labs opens it, because it produces the same family of outputs, point clouds and Gaussian splats, from fewer inputs (Dealroom).

Interactive world models stream frames. You open a session, send control inputs, and receive a video stream at 20 to 24 frames per second. The API is a socket, not a request. Among the models compared here, Odyssey-2 is the one that exposes such an API publicly; Genie 3 and GWM Worlds 2 run the same kind of loop but only inside their own products.

Video generation models return a file. Prompt in, clip out, priced per second. Sora 2 and MiniMax H3 belong here. They are often called world models in coverage because they simulate physics convincingly, but they have no camera-path input and no 3D output, which is the distinction World Labs draws for Atlas (SiliconANGLE).

Knowing which shape you need removes most of the table before you compare prices.

The APIs you can use today

Marble World API (World Labs)

The World API opened on January 21, 2026 (World Labs). It accepts text, a single image, multiple images, a 360° panorama or a video, and returns a navigable Gaussian-splat world. Since April 2, 2026 the default model is Marble 1.1, with Marble 1.1 Plus for worlds that expand beyond one space (Radiance Fields).

• Endpoint shape. POST /marble/v1/worlds:generate returns an operation you poll; GET /marble/v1/credits reports balance (World Labs docs).
• Exports. Gaussian splats as .spz or .ply and a mesh as GLB; the API does not export video, though the Marble web app can record a fly-through. PLY and SPZ export is free, a high-quality mesh is 3,500 credits or about $2.80 (World Labs docs).
• Price. $1 buys 1,250 credits with a $5 minimum; a Marble 1.1 world costs 1,500 credits, about $1.20, and a draft world 150, about $0.12 (World Labs docs). Marble 1.1 Plus adds a variable component of up to 1,500 credits on top of the base, reported as 300 credits per extra dynamic cube (Radiance Fields).

It is the only public API in this list that returns 3D geometry you can export and reuse; the others return video. Walkthroughs are in the Marble World API guide and the TypeScript quickstart; the credit maths is on the pricing page.

Odyssey-2 developer API

Odyssey announced Odyssey-2 on January 23, 2026 with a public product and a developer API at developer.odyssey.ml (Odyssey). Odyssey-2 Pro streams 720p at 22 frames per second with roughly 50 ms response time and supports minutes-long continuous interaction (Air Street Press). There is no 3D export; the product is the stream. If your application is a live, explorable experience rather than an asset pipeline, this is the only public API for it among the models compared on this page; other real-time video vendors exist outside this comparison. We have not independently verified Odyssey’s pricing and do not quote it here.

MiniMax H3 (Hailuo 3.0)

MiniMax released H3 on July 31, 2026 as a 33-billion-parameter omni-modal video model with open weights, hosted on fal, ComfyUI and the MiniMax API (fal, MarkTechPost). It generates 4 to 15 second clips at native 2K and 24 frames per second with 32 kHz stereo audio and lip-synced dialogue. It is a video model rather than a world model in the geometric sense, but for teams that were using Sora 2’s API it is the most direct replacement with a public endpoint. The full feature comparison is in Atlas vs MiniMax H3.

Sora 2: read before you integrate

Sora 2’s API is priced at $0.10 per second for 720p and $0.30 to $0.50 per second for sora-2-pro, with batch at half price (eesel). It also stops working on September 24, 2026: OpenAI discontinued the Sora app on April 26, 2026 and announced the API shutdown for September (OpenAI Help). Do not start a new integration against it. It is in the table because people still search for it and because its per-second pricing is the reference point most developers have for what video generation costs.

The world models with no public API

Atlas (World Labs)

Atlas was announced on September 1, 2026 and is available only to early-access partners through a request form. No documentation, endpoints or pricing have been published (Implicator). Because Atlas “builds on the Marble model” and World Labs already runs the World API (Dealroom), the likeliest form of an Atlas API is a new model identifier on the same platform. That is why the practical advice is to build on Marble now and keep the model name behind an interface. Status is tracked on the Atlas World Model API page, and the request process is described in how to get Atlas early access.

Genie 3 (Google DeepMind)

Genie 3 renders a playable 720p world at 24 frames per second and keeps it consistent for several minutes (DeepMind). Since January 29, 2026 it has been reachable as Project Genie for Google AI Ultra subscribers in the US (Wikipedia). That is a consumer surface, not an API: there is no endpoint, key or pricing for developers. See Atlas vs Genie 3 for the feature comparison.

GWM Worlds 2 (Runway)

Runway announced GWM Worlds 2 on September 3, 2026 as a research preview. The launch page routes business enquiries through a contact form; there is no self-serve access, no published price and no API (Runway). The detail is in GWM Worlds 2 pricing and access.

Comparing prices when the units differ

The three public APIs do not share a unit, so “cheapest” needs a use case.

• Per world. Marble charges a flat 1,500 credits, about $1.20, for a scene you can render from any angle indefinitely and export for free as PLY. Cost does not scale with how long users spend inside it.
• Per second. Sora-style pricing at $0.10 per second means $1.20 buys 12 seconds of one fixed camera move. A second angle is a second charge.
• Per session. Streaming world models bill for compute while the session runs. Odyssey’s rates are not quoted here; assume a live stream costs more per minute of user time than either of the above, because the GPU is reserved for one user in real time.

For asset pipelines, per-world pricing wins by a wide margin. For a 15-second marketing clip, per-second pricing is fine. For a live experience there is no cheap option yet.

What to check before you commit to a world model API

1. Output format. If anything downstream needs geometry, only Marble qualifies today. Preview a sample export in the free splat viewer before you build around it.
2. Latency model. Marble jobs take minutes; do not block a web request on them. Streaming APIs need a persistent connection and a client that can render at 20+ frames per second.
3. Camera control. Only Atlas takes an explicit camera path, and it is not public. Marble gives you a static world you fly a camera through yourself, which is often the better trade for repeatable shots.
4. Shutdown risk. Sora 2 is the reminder that a hosted API can disappear. Open-weights options such as MiniMax H3 remove that risk at the cost of running inference yourself.
5. Terms for generated content. Commercial rights differ by vendor and plan. Check them before shipping outputs.

What changes next

Two events would redraw the table. The first is an Atlas endpoint on the World API, which would make World Labs the only vendor with both a shipping 3D API and a camera-controlled video API. The second is developer access to Genie 3 or GWM Worlds 2, which would give the streaming category a second and third public option. Neither has a date. This site logs each change in the changelog within 48 hours, and the comparison hub holds the side-by-side feature tables behind every row above.