World models
We track 34 benchmarks in world models, or 37 if you count each version. They run from everyday level to expert level.
OpenAI held the top world models score 95% of the contested time
That's about 44 of the 46 months since January 2023, averaged over the benchmarks below. The easiest benchmarks are at the top. Versions of one test share a row, with a dark tick at each new version. We stop the clock, and fade the rest of the row, when the best score passes 85%, or at the newest score if 45 days go by without another.
OpenAI Google DeepMind
Share of contested time at number one
OpenAI 95% 
Google DeepMind 5%
Google DeepMind 5% Scores come from each benchmark’s own leaderboard, from Epoch AI’s runs and from labs’ reports on their own models. We date Epoch runs by the model’s release and lab reports by when they came out. Labs pick their own settings, so a lab’s claim can beat an independent run of the same model. Turn off lab claims to see independent runs only.
OpenAI Google DeepMind
Share of contested time at number one, independent runs only
OpenAI 95% Google DeepMind 5%
Google DeepMind 5% Only each benchmark’s own leaderboard and Epoch AI’s runs, dated by the model’s release. We leave out benchmarks that only labs report.
From everyday to expert level
Grouped by who could solve the tasks, from things most adults can do to open research problems. The bar under each name is its best score.
Everyday
Most adults, without training
Expert
Specialists with years of training, often over hours of work
Another 14 have no level. Levels don’t apply to arenas, safety evaluations or composite indices.
By kind
Video world models
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| RBench 650 image+text prompts for robot video generation (tasks and embodiments). MLLM judge scores task completion, physical plausibility, stability. | Graded by a model |  60.7% | Not in cards | |
| VGI-Bench Video generation models solve 27 visual reasoning tasks with 810 instances through generated frames. Scoring checks that the process is valid, not only the final frame. | Everyday | Graded by a model |  51% | Not in cards |
| PAI-Bench Physical-AI video generation from 1,044 prompts in six domains, VLM-QA plus quality metrics. Companion control and understanding tracks. | Graded by a model | 82.1% | Not in cards | |
| 1X World Model Challenge Predict future egocentric frames of 1X's EVE humanoid from video and state. Compression scored by token cross-entropy, Sampling by PSNR on private tests. | Checked automatically | Not in cards | ||
| DreamGen Bench Video generation benchmark measuring how well video models adapt to a target robot embodiment. Part of NVIDIA's DreamGen pipeline. | Graded by a model | Not in cards | ||
| EWMBench Embodied world models generate robot manipulation videos from language. Scored on scene consistency, motion correctness, semantic alignment. | Checked automatically | Not in cards | ||
| WorldArena Embodied world models generate RoboTwin 2.0 task videos from first frame plus text or actions. 15 video metrics (EWMScore) plus utility tracks. | Checked automatically | Not in cards | ||
| WorldExam Controllable video generators are tested as world models. Levels run from visual appearance to whether the depicted world reacts plausibly beyond what the prompt states. | Everyday | Graded by a model | Not in cards | |
| WorldModelBench Video generators judged as world models on instruction following and physics adherence in application domains. Human labels and a fine-tuned judge. | Graded by a model | Not in cards | ||
| WorldScore 3,000 next-scene generation tasks with camera-trajectory layouts. Scores 3D/4D and video world generators on controllability, quality, dynamics. | Checked automatically | Not in cards | ||
| WorldSimBench Dual evaluation of video world simulators: human-preference perceptual scoring plus action-level embodied tasks in games, driving, manipulation. | Graded by a model | Not in cards |
Interactive world models
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| HarnessEval-W Video and world models generate rollouts from image and action prompts. An evaluator agent plans checks, spawns sub-agents, gathers spatiotemporal evidence and scores observation, transition and persistence with written reasoning. | Expert | Graded by a model | 75.5 | Not in cards |
| QwenWorldBench Models act as world models that simulate agent environments in 7 domains, such as terminal, web, and Android. A rubric judge scores five dimensions. | Graded by a model |  62.1% | Fading | |
| WBench Interactive video world models run 289 multi-turn cases with navigation, subject actions, event edits and perspective switches. 22 metrics cover quality, adherence, consistency and physics. | Everyday | Graded by a model |  79.0 | Not in cards |
| AgentWorldBench Language world models predict the next environment observation after an agent action, in 7 domains from MCP and search to Android and desktop OS. A judge model scores each prediction against the real observation on five dimensions. | University | Graded by a model | 56.4 | In use |
| Dual Exam Multimodal agents are tested under fixed environment responses, and world models under fixed actions. The agent track has 1,088 items and the world model track 400 cases. | Everyday | Checked automatically | Not in cards | |
| GlobeBench Models act as text world models. Given an agent's past actions and observations, they predict the environment's next observation across code, enterprise, games, math and science, and dialogue. | Expert | Graded by a model | Not in cards | |
| HappyWorld-Bench World models are tested on the quality of the worlds they generate and on staying consistent and responsive while agents explore, interact and modify them. | Everyday | Graded by a model | Not in cards | |
| PROWBench Programmable world models must render video that follows the rules and interactions their program specifies. The test checks visual fidelity to fine-grained rules. | Everyday | Checked automatically | Not in cards | |
| World-in-World Closed-loop platform: world models plan for embodied agents in four environments via a standard action API. Scored by task success. | Everyday | Checked automatically | Not in cards | |
| WorldArena 2.0 Extends WorldArena to visuotactile prediction, world models as RL environments, and simulated plus real multi-embodiment robot platforms. | Checked automatically | Not in cards | ||
| WorldModelGym World models predict returns for five candidate action sequences across 100+ Atari, Meta-World, DMC and classic-control tracks. Scored by normalized regret. | Checked automatically | Not in cards |
Physical prediction
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| VPCT Models view an image of a ball, ramps, and three buckets and predict which bucket the ball falls into. Scored by accuracy on 100 problems. | Everyday | Checked automatically | 91% | Fading |
| Physics-IQ Verified Video models continue 198 real filmed physics experiments from their first frames. Generated videos are compared with the real outcome, with every sample and metric weighted equally. | Everyday | Checked automatically | 10.9 | Not in cards |
| ContPhy Models infer physical properties such as mass and density of fluids and soft bodies from videos and predict their dynamics. | Everyday | Checked automatically | 67.4% | Fading |
| MVPBench Models answer video questions on spatial, temporal and intuitive physics. Each question comes in a pair of near-identical videos with opposite answers, so text or single-frame shortcuts fail. | Everyday | Checked automatically | 72.5% | Dropped |
| IntPhys 2 Models watch simulated videos and judge whether each event is physically possible, testing object permanence, immutability, continuity and solidity. Scored by accuracy. | Everyday | Checked automatically |  68.3% | In use |
| Physics-IQ Video models continue 396 filmed physical events from a frame or 3 s clip. Motion compared with the real continuation. | Checked automatically | 46.4% | Not in cards | |
| VideoPhy-2 200 action-centric prompts. Generated videos human-rated for semantic adherence, physical commonsense and physical-rule grounding. | Graded by people | 23.3% | Not in cards | |
| 4DCodeBench Agents reconstruct dynamic scenes from video as executable graphics programs, including physical simulations, so rendered output matches the observed motion. | Expert | Checked automatically | Not in cards | |
| CausalVQA Models answer questions about real egocentric videos that probe physical cause and effect, in counterfactual, hypothetical, anticipation, planning and descriptive types. | Everyday | Checked automatically | Not in cards | |
| ExeWorldBench Coding agents build executable code-based 3D worlds that capture how a world works, testing world understanding and its realization in code. | Expert | Checked automatically | Not in cards | |
| PAWBench Video generators are run many times on the same start and action, and the outcomes become a distribution over physical behaviors. The test checks whether it matches reference probabilities and covers every valid behavior. | Everyday | Checked automatically | Not in cards | |
| PhysicalRealismBench-U VLM watches 1,300 synthetic Isaac Sim videos, flags physics-law violations, names object and frames. Scored by joint realism F1. | Everyday | Checked automatically | Not in cards | |
| PhysVista Vision-language models perceive, reason about and judge whether videos follow real-world physics. The benchmark also tests if they can judge the physical realism of generated videos. | Everyday | Checked automatically | Not in cards | |
| VideoPhy Text-to-video models generate clips from captions of physical interactions between solids and fluids. Human raters judge caption match and physical commonsense. | Everyday | Graded by people | Not in cards | |
| YoCausal Video generation models are tested for causal and arrow-of-time understanding by comparing real videos with their time-reversed versions, in two levels. | Everyday | Checked automatically | Not in cards |