The current Physics-IQ Verified leaderboard backs the top-line ranking, but the details matter. As of Oct. 6, the No. 1 entry is not the plain model alone. It combines FLUX 3 [large] with Claude Opus 5.5-generated prompts and a best-of-eight selection system using WMReward and candidate consensus. That configuration scores 64.35 ± 0.22 in the multiframe video-to-video category.
A FLUX 3 entry using the same prompt setup without the best-of-eight selector ranks second overall at 61.11 ± 0.55. The model family also occupies the top two image-to-video positions, where the augmented and non-augmented entries score 54.70 ± 0.41 and 51.11 ± 0.44, respectively.
That makes Black Forest Labs' claim that “FLUX 3 is #1” accurate for its highest-scoring configuration. The leaderboard also supports a narrower point: its plain FLUX 3 entry sits above the other listed entries that do not use a selection layer.
How Physics-IQ tests video models
Physics-IQ uses real camera footage rather than rendered scenes. Its scenarios cover collisions, fluid dynamics, gravity, material properties, light, shadows, magnetism and other physical phenomena. Each scenario is filmed from three perspectives and recorded twice to capture natural variation.
For image-to-video tests, a model receives a switch frame and generates what comes next. Multiframe video-to-video systems receive a short conditioning clip. A perfect score of 100 would represent videos that differ from the real continuation only through physically plausible randomness while obeying the principles being tested.
Physics-IQ Verified is the recommended variant. Its maintainers describe it as an update with improved prompts, masks and scoring. The submission workflow recommends four runs and reports a mean plus standard deviation, which makes the uncertainty around each score visible.
Black Forest Labs' initial submission record says each of its four runs contained 198 five-second clips at 24 frames per second and 832-by-480 resolution. The issue was opened Oct. 4 and closed after the entries were merged Oct. 5. The leaderboard has since been updated with higher scores tied to the Claude Opus 5.5 prompt configuration.
The selection layer is part of the story
The best-of-eight system generates eight candidates, then ranks them using WMReward, described in the submission as a V-JEPA 2 “surprise” signal, and a consensus score based on how closely each candidate agrees with the others on Physics-IQ's metrics. It selects the candidate with the best combined rank.
The submission also discloses that the team chose its sampling settings and prompt variant through ablations scored on Physics-IQ Verified. The selector itself was picked from 10 pre-declared rules based on performance over the same candidate pools. Those choices do not erase the result, but they narrow its meaning: the top score reflects a tuned generation-and-selection pipeline, not just one untouched model output.
Black Forest Labs argues that the underlying physics capability benefits both FLUX 3 Video's motion generation and FLUX 3 Action's robotics predictions. Physics-IQ alone cannot establish those broader product claims. It measures performance on a defined set of short filmed experiments, not general physical reasoning, robotics safety or reliability in the open world.
There is one more attribution wrinkle. The repository states that Physics-IQ Verified is an independent third-party benchmark and is not endorsed or verified by Google DeepMind. The Oct. 4 submission also described FLUX 3 [large] as unreleased. The ranking therefore describes a submitted evaluation configuration; it does not establish outside access to the same setup.
