You’ve watched the demos. Sora generates a video of a mechanical arm pouring coffee, and the crowd goes wild. The pixels are pristine. The lighting is perfect. The water looks genuinely wet.
But here’s what nobody at those keynotes tells you: when you ask that same model to predict what happens when the cup tips over — where the liquid flows, how fast it spills, whether the handle snaps — it fails. Badly.
Looking real and being real are not the same thing. And in the race to build world models, we’ve been measuring the wrong one.
This isn’t a minor quibble. It’s the single most important distinction in AI right now, and almost nobody is talking about it because the marketing machines at OpenAI, Google, and NVIDIA are drowning out the signal.
At WAIC 2026 in Shanghai this July, I walked the exhibition floor and watched something revealing. Robot companies showed off humanoid mechs and dexterous arms. Next door, video generation companies played stunning demos on wall-to-wall screens — mechanical arms pouring water, drones weaving through forests, virtual characters turning around. Every booth’s presenter said the same two words: world model.
But when investors walked between booths asking “how good is your model?” they got completely different answers. Someone would pull up a VBench score. Someone else would show a robotics task success curve. Someone else would cite sim-to-real transfer data. Same word, totally different yardsticks.
That confusion isn’t accidental. It’s convenient.
Four Benchmarks, Four Truths
Right now, the “world model” space is a jigsaw puzzle with four pieces, and each piece has its own leaderboard:
VBench answers “does it look real?” — measuring visual fidelity, motion smoothness, aesthetic quality. This is where Sora and Veo shine. On VBench 2.0, Google’s Veo 3 sits at 66.72%, and the top models cluster tightly between 84-88% on the original scale. The gap is closing. Everyone looks good here.
WorldModelBench answers “does it understand?” — testing whether models follow physics and common sense across robotics, driving, industrial, and human activity domains. Even the best model only correctly completes 61% of instructions. 12% of generated videos violate conservation of mass. 11% show objects passing through each other like ghosts.
RoboTwin 2.0 answers “can it act?” — running 50 dual-arm manipulation tasks in simulation, from tool use to deformable object handling. In hard settings, most methods still succeed less than 20% of the time. Translation: robots can do specific tricks but can’t generalize.
And then there’s the fourth benchmark — the one that actually matters.
Physics-IQ: The Benchmark Nobody Hypes
When you push a glass across a table, your brain predicts how far it slides and whether it tips. When you tilt a kettle, you know where the water comes out. When you push two magnets together, you feel attraction or repulsion before contact.
Robots need this exact ability. They must look at the current state of the world and predict what happens next — then act on that prediction. If the prediction is wrong, everything downstream collapses.
Google DeepMind and INSAIT built a benchmark for this called Physics-IQ. They filmed 66 real physical experiments — solid mechanics, fluids, optics, thermodynamics, magnetism — with three camera angles, two takes each, producing 396 eight-second clips. Models see the first three seconds and must predict the next five. The score is normalized against natural variation: even reality doesn’t reproduce identically when you film it twice.
This is the bedrock. And what it reveals is devastating to the conventional narrative.
The model that best understands physics isn’t from OpenAI, Google, or NVIDIA. It’s from a Chinese startup most people in Silicon Valley have never heard of.
Sand.ai’s Magi-1 scored 56.0% on Physics-IQ when it topped the leaderboard in April 2025. The previous leader, Google’s VideoPoet, sat at 29.5%. Magi-1 nearly doubled it. It held the #1 spot for roughly 13 months, leaving only briefly.
NVIDIA’s flagship Cosmos3-Super, boosted with WMReward, surged to 63.4% in May 2026 — a three-week reign. Then a Magi-1-based enhanced system reclaimed the top with 64.5%.
On the corrected Physics-IQ Verified leaderboard published in June 2026, Magi-1’s 24B enhanced version scored 58.2 (first place), its bare model scored 48.4 (second place), Cosmos3-Super scored 39.5, and Sora 2 scored just 26.5.
Let that sink in. Sora 2 — the model that generated a million retweets, dominated tech headlines, and became synonymous with “AI video” — scores less than half of the leader on the benchmark that actually tests whether the model understands physical reality.
Why Magi-1 Wins: A Fundamentally Different Bet
The answer isn’t more data or bigger GPUs. It’s architecture.
Sora, Kling, and most mainstream video models use diffusion: they generate an entire video at once, denoising all frames simultaneously with no strict temporal ordering. It’s like painting a mural — every pixel appears together.
Magi-1 takes a completely different approach. It generates video autoregressively, block by block, in strict chronological order — 24 frames per block. Within each block, diffusion handles visual quality. Between blocks, autoregressive chaining ensures that what comes next depends on what came before. Past frames are written and locked; future frames build on them.
This is exactly how language models work: predict the next token based on everything before it. Magi-1 just does it with video blocks instead of text tokens.
Diffusion models generate beautiful lies. Autoregressive models make predictions. Only one of those is a world model.
The architecture enables streaming generation — video plays as it’s being created — and theoretically infinite continuation. For world models, that means simulation can run indefinitely, and prompts can change mid-stream to redirect the scenario. You can’t do that with diffusion.
There’s a deeper philosophical alignment here too. Yann LeCun has argued for years that AI needs internal world models that predict, not just generate. His JEPA architecture — I-JEPA, V-JEPA, V-JEPA 2 — predicts in representation space rather than pixel space, deliberately ignoring unpredictable details. Magi-1 predicts in pixel space block by block; JEPA predicts in abstract representation space. Different implementations, same conviction: a world model must predict what happens next.
The Coordinate System Is Shifting
For the past two years, video model launches competed on “does it look real?” — resolution, duration, aesthetic scores. Physics-IQ changes the question to “does it predict physical reality accurately?”
This shift won’t appear on any product launch invitation. No CEO will announce “our model still can’t tell when a cup will tip over.” But it will determine where capital flows next year, where talent migrates, and which companies survive when the hype fades.
Because here’s the uncomfortable truth: a model that generates photorealistic video of a robot pouring water but can’t predict whether the cup is full or empty is not a world model. It’s a very expensive screensaver.
The benchmarks will iterate. The rankings will shift. Cosmos will improve. Sora will iterate. New challengers will emerge. But the principle is now locked in: if your model can’t predict physics, it doesn’t understand the world — no matter how pretty its output looks.
Once the right yardstick exists, you can’t un-measure things. The models that score poorly on physics prediction will either adapt or die. And the companies still selling pixel-quality as proof of world understanding are selling you a story, not a product.
The race for world models didn’t end with Sora’s launch. It barely started. And right now, the leader isn’t who you think.
FAQ
Q: Isn't Sora still improving? Won't it catch up on physics prediction eventually?
A: Maybe, but the architecture matters. Diffusion generates all frames simultaneously with no strict temporal causality — it's fundamentally not built for sequential prediction. Magi-1's autoregressive block generation is architecturally closer to how prediction actually works. Sora would need to rethink its core design, not just add more training data.
Q: What does this mean for investors and builders evaluating world model companies?
A: Stop looking at VBench scores and demo reels. Ask companies for their Physics-IQ results. If they can't answer or dodge the question, that tells you everything. The companies that matter are the ones benchmarking against physical prediction, not pixel aesthetics.
Q: Is this just Chinese AI hype replacing Silicon Valley hype?
A: No — it's a Chinese startup winning on a benchmark designed by Google DeepMind. The yardstick is Western-built; the winner just happens to be Chinese. That's what makes it uncomfortable for the conventional narrative. The architecture choice, not nationality, explains the win.