Best world model API
4 models · updated 2026-07-15
The verdict
World Labs Marble leads — 3 of 4 models rank World Labs Marble the top pick.
Not unanimous: Claude picks NVIDIA Cosmos.
As of 2026-07-15, ChatGPT, Claude, Gemini and Grok collectively rank World Labs Marble #1 for world model api on ModelsAgree by aggregate score. The models' case: Best overall developer option: generates persistent, navigable 3D worlds from text, images, panoramas, video, or multi-view inputs, with web rendering and exportable. The models' main caveat: Gaussian-splat-style worlds lack dependable object physics and game-engine-grade geometry. The strongest alternative is NVIDIA Cosmos — The most complete world-model stack a practitioner can actually build on today — Predict/Transfer/Reason world foundation models with open weights. Not unanimous: Claude picks NVIDIA Cosmos. Source: https://modelsagree.com/best/best-world-model-api (modelsagree.com, CC BY 4.0).
Combined ranking
- 1GPT #1Claude #2Gemini #1Grok #1
Best overall developer option: generates persistent, navigable 3D worlds from text, images, panoramas, video, or multi-view inputs, with web rendering and exportable spatial assets
+ model takes & fixes− hide details
GPT Best overall developer option: generates persistent, navigable 3D worlds from text, images, panoramas, video, or multi-view inputs, with web rendering and exportable spatial assets
Gemini Leads in spatial intelligence by generating persistent, geometrically consistent 3D environments from text, image, or video inputs, enabling programmatic integration into games and simulations.
Grok Leading public API for generating persistent, navigable, high-fidelity 3D worlds from text/images/video/panoramas; strong spatial consistency, editability, and browser-based exploration make it highly practical for developers/creators building interactive apps/simulations; multimodal strengths and real-world usability edge it out.
Claude The first commercially shipping generative-3D-environment product that fits real pipelines — text/image/video prompts to persistent, explorable Gaussian-splat worlds with mesh/splat export into Unity, Unreal, and web viewers, plus editing and expansion tools; for anyone who needs an actual 3D scene artifact rather than a video, nothing else commercial comes close.
Where it falls shortper GPT Gaussian-splat-style worlds lack dependable object physics and game-engine-grade geometry
per Claude Worlds are static geometry — no dynamics, physics, or agent interaction — and programmatic/API access is much less mature than the web app, so it's a scene generator, not an interactive simulator.
per Gemini Operating as a proprietary, credit-based commercial API makes it expensive and dependent on cloud connectivity.
per Grok Commercial pricing and potential compute costs limit heavy local/self-hosted use (not ideal for fully open-source or low-budget offline workflows).
- 2GPT #3Claude #1Gemini #2Grok #3
The most complete world-model stack a practitioner can actually build on today — Predict/Transfer/Reason world foundation models with open weights plus hosted NIM APIs, physics-aware video generation with camera and action conditioning, and tight Omniverse/Isaac integration for 3D simulation pipelines; it's the de facto choice for robotics, AV, and synthetic-data teams, which is where world models deliver real economic value in 2026.
+ model takes & fixes− hide details
Claude The most complete world-model stack a practitioner can actually build on today — Predict/Transfer/Reason world foundation models with open weights plus hosted NIM APIs, physics-aware video generation with camera and action conditioning, and tight Omniverse/Isaac integration for 3D simulation pipelines; it's the de facto choice for robotics, AV, and synthetic-data teams, which is where world models deliver real economic value in 2026.
Gemini Offers an open-weights Mixture-of-Transformers (MoT) foundation model for physics-grounded world simulation and video generation with deep integration in the NVIDIA Omniverse ecosystem.
GPT Strongest customizable physical-AI stack, combining deployable APIs, open-weight world foundation models, video-to-world generation, post-training tools, and NVIDIA’s simulation ecosystem
Grok Open-weight world foundation models with superior physics/reasoning for robotics/AV/physical AI; flexible APIs, NIMs, and tools for custom post-training/simulation; unmatched for action-conditioned, real-world grounded interactive video/3D.
Where it falls shortper GPT Infrastructure-heavy and aimed mainly at robotics or autonomous-vehicle teams, not quick creative-world generation
per Claude Built for physical-AI simulation, not creative or consumer interactive worlds — expect serious GPU and integration lift, and it won't hand you a playable game-like environment out of the box.
per Gemini Requires significant local enterprise GPU infrastructure to run the larger parameters (up to 64B) locally.
per Grok Heavily optimized for enterprise/physical AI (requires significant GPU/resources); less turnkey for general creative 3D storytelling or lightweight web apps.
- 3GPT #2Claude #4Gemini #3Grok —
Best near-tie for real-time interactive video, producing controllable closed-loop photorealistic environments especially valuable for autonomous-driving and embodied-agent training
+ model takes & fixes− hide details
GPT Best near-tie for real-time interactive video, producing controllable closed-loop photorealistic environments especially valuable for autonomous-driving and embodied-agent training
Gemini Provides a dedicated gRPC SDK for real-time action-conditioned simulation, producing synchronized multi-camera perspectives for closed-loop autonomous training.
Claude The only production API delivering genuinely real-time interactive generative video — MirageLSD transforms live streams frame-by-frame at ~25fps with sub-100ms latency and unlimited duration, and the Oasis line demonstrated fully playable model-generated worlds; near-tie with Sora 2 on this list, ranked below it only because Decart's fidelity and generality are narrower.
Where it falls shortper GPT Primarily a visual simulator, with weaker long-horizon consistency and explicit physical state than conventional simulation
per Claude Focused on real-time style transformation and game-like demos at modest resolution — it won't generate arbitrary high-fidelity environments or exportable 3D geometry.
per Gemini High API billing rates (~$0.02/second) and narrow specialization for robotics and autonomous driving navigation.
- 4GPT —Claude —Gemini —Grok #2
Exceptional real-time interactive environment generation at 720p/24fps with strong consistency over minutes; text-to-dynamic-worlds excels for diverse, explorable scenarios and research/education use cases; powers robust navigation and agent training.
+ model takes & fixes− hide details
Grok Exceptional real-time interactive environment generation at 720p/24fps with strong consistency over minutes; text-to-dynamic-worlds excels for diverse, explorable scenarios and research/education use cases; powers robust navigation and agent training.
Where it falls shortper Grok Primarily research preview/limited access (e.g., Ultra subscribers or cohorts) rather than fully open public API, restricting broad production deployment.
- 5GPT —Claude #3Gemini —Grok —
The strongest physics-consistent video world simulation available through a plain, well-documented REST API (OpenAI platform and Azure), with synced audio, a pro quality tier, and the broadest ecosystem of SDKs and tooling; if "world model" means generating believable video of how a scene evolves, this is the most accessible high-end option.
+ model takes & fixes− hide details
Claude The strongest physics-consistent video world simulation available through a plain, well-documented REST API (OpenAI platform and Azure), with synced audio, a pro quality tier, and the broadest ecosystem of SDKs and tooling; if "world model" means generating believable video of how a scene evolves, this is the most accessible high-end option.
Where it falls shortper Claude Output is passive video only — no interactivity, no 3D assets, short clips at meaningful per-second cost — so it's a renderer of worlds, not an environment you can enter.
- 6GPT —Claude —Gemini #4Grok —
Delivers stable, browser-embeddable interactive 3D scenes and video generation (Ray 3.2) with granular keyframe controls and engine plugins for creative workflows.
+ model takes & fixes− hide details
Gemini Delivers stable, browser-embeddable interactive 3D scenes and video generation (Ray 3.2) with granular keyframe controls and engine plugins for creative workflows.
Where it falls shortper Gemini Lacks true action-conditioned physics simulation, acting more as a visual generator than a dynamic world simulator.
- 7GPT —Claude —Gemini —Grok #4
Strong causal next-state interactive video/world model with real-time streaming API; good physics/dynamics and developer-friendly endpoints for online/offline workflows; solid general-purpose value for practitioners building interactive experiences.
+ model takes & fixes− hide details
Grok Strong causal next-state interactive video/world model with real-time streaming API; good physics/dynamics and developer-friendly endpoints for online/offline workflows; solid general-purpose value for practitioners building interactive experiences.
Where it falls shortper Grok Newer entrant with potentially less proven long-horizon consistency or ecosystem maturity compared to leaders (not for ultra-demanding physics-critical robotics).
- 8GPT #4Claude —Gemini —Grok —
Most production-ready creative-video choice, with strong visual fidelity, mature tooling, scalable inference, and GWM-derived real-time character capabilities
+ model takes & fixes− hide details
GPT Most production-ready creative-video choice, with strong visual fidelity, mature tooling, scalable inference, and GWM-derived real-time character capabilities
Where it falls shortper GPT Its general explorable-world technology is not exposed as a broad self-service environment API
- 9GPT #5Claude —Gemini —Grok —
Excellent programmable video generation with native audio, strong temporal coherence, reference-image control, endpoint-frame control, and clip extension
+ model takes & fixes− hide details
GPT Excellent programmable video generation with native audio, strong temporal coherence, reference-image control, endpoint-frame control, and clip extension
Where it falls shortper GPT Generates linear clips rather than action-conditioned interactive worlds
- 10GPT —Claude —Gemini —Grok #5
Fully open-source with excellent real-time interactivity (low-latency 720p/60fps, action controls like WASD/attacks), long-horizon stability, and self-correcting physics from game data; highly accessible for customization, gaming, and embodied AI research.
+ model takes & fixes− hide details
Grok Fully open-source with excellent real-time interactivity (low-latency 720p/60fps, action controls like WASD/attacks), long-horizon stability, and self-correcting physics from game data; highly accessible for customization, gaming, and embodied AI research.
Where it falls shortper Grok Open-source deployment demands local compute/expertise; may trail closed models in photorealism or seamless multimodal inputs for non-technical users.
- 11GPT —Claude #5Gemini —Grok —
The best open-source path to 3D world generation — HunyuanWorld 1.0 plus Voyager turn text or images into explorable, mesh-exportable 360° scenes with camera-controllable world video, self-hostable under a permissive license at zero API cost; near-tie with Decart, and it wins the slot over closed previews because practitioners can actually run it.
+ model takes & fixes− hide details
Claude The best open-source path to 3D world generation — HunyuanWorld 1.0 plus Voyager turn text or images into explorable, mesh-exportable 360° scenes with camera-controllable world video, self-hostable under a permissive license at zero API cost; near-tie with Decart, and it wins the slot over closed previews because practitioners can actually run it.
Where it falls shortper Claude Research-grade output — panorama-anchored scenes with limited free-roam range and rough geometry — and there's no managed API, so you own the GPU hosting and integration work.
Just missed the top 5
GPT Google DeepMind Genie 3 — exceptional interactive-world capability but Project Genie is a constrained research product, not a general developer API · Odyssey — promising real-time interactive video research, but insufficiently mature and broadly accessible for typical production use
Claude Google DeepMind Genie 3 — technically the most impressive interactive world model — real-time playable 720p worlds with minutes-long consistency — but still a gated research preview with no public API as of early 2026, so its practitioner value is near zero
Gemini Google DeepMind Genie 3 — lacks a public developer API, being restricted to Google Labs' Project Genie web interface · LingBot-World 2.0 — lacks a managed, cloud-hosted API service, requiring developers to self-host the complex simulation stack
Grok Runway GWM-1 — strong robotics/world simulation via SDK but more specialized/less mature public API accessibility for general 3D interactive environments
By model
ChatGPT
- 1.World Labs Marble
- 2.Decart Oasis 3
- 3.NVIDIA Cosmos
- 4.Runway
- 5.Google Veo 3.1
Claude
- 1.NVIDIA Cosmos
- 2.World Labs Marble
- 3.OpenAI Sora 2
- 4.Decart Oasis 3
- 5.Tencent HunyuanWorld
Gemini
- 1.World Labs Marble
- 2.NVIDIA Cosmos
- 3.Decart Oasis 3
- 4.Luma
Grok
- 1.World Labs Marble
- 2.Google DeepMind Genie 3
- 3.NVIDIA Cosmos
- 4.Odyssey-2 Pro
- 5.LingBot-World 2.0
Common questions
What is the best world model api according to AI models?
World Labs Marble leads. 3 of 4 models rank World Labs Marble the top pick. The current top 3: World Labs Marble, NVIDIA Cosmos, Decart Oasis 3. Ranked by asking ChatGPT, Claude, Gemini, Grok the same buying question and merging their top-5 picks, updated 2026-07-15. Source: modelsagree.com.
Which world model api did each AI model pick first?
ChatGPT: World Labs Marble. Claude: NVIDIA Cosmos. Gemini: World Labs Marble. Grok: World Labs Marble.
Do the AI models agree on the best world model api?
Not unanimous. Claude picks NVIDIA Cosmos.
How is this world model api ranking made?
ChatGPT, Claude, Gemini, Grok are each asked the same buying question in a fresh session with no system steering. Their top-5 answers are merged (rank 1 = 5 pts … rank 5 = 1 pt) into the consensus ranking, re-polled on demand and tracked over time.
More on how polling works: full methodology →
Cite this ranking
ModelsAgree, “Best world model API” — merged ranking from ChatGPT, Claude, Gemini & Grok, polled 2026-07-15. https://modelsagree.com/best/best-world-model-api (CC BY 4.0)
Tracked by ModelsAgree · rank 1 = 5 pts … rank 5 = 1 pt · re-polled on demand