Observed facts: Artificial Analysis reports a v1.0 evaluation of 21 models across 55 written briefs. Each model works in an offline sandbox and returns an index.html scene; the run is capped at three hours or 250 turns. Human raters compare pairs of scenes for the same brief without knowing their authors, and the ranking uses Elo. Claude Sonnet 5.5 leads at 1194, while the top four are described as statistically tied because their 95% confidence intervals overlap.
Mechanism: The shared brief and delivery contract create comparable outputs, while blinded pairwise votes capture human preference across whole scenes. Why now: As models are asked to build usable interactive experiences, a benchmark needs to test the artifact people encounter, not only the intermediate code. First implication: Teams evaluating generative 3D workflows can use task-specific preference evidence to compare systems, while treating the ranking as bounded to this benchmark.
The article also gives a single-run cost example, but neither that sample nor the ranking establishes broad cost efficiency, reliable performance across all 3D work, or production readiness. Evidence boundary: AA-3D reports results from its own 55-brief v1.0 setup; these are not a universal measure of model quality.
First watch condition: Watch whether later benchmark versions expand the brief set and report uncertainty or repeatability in enough detail to show whether rank differences persist. With no earlier New Runtime record supplied, this starts a distinct analytical record on artifact-level preference evaluation for interactive 3D generation.