Imagine you asked sixteen professional chefs to cook dinner for eighty different occasions — a child's birthday, a funeral reception, a state dinner, a beach picnic — across two cuisines each. Then you scored them not just on whether the food was edible, but on three separate axes: did they all make the same dish (originality), did they actually change their approach when the occasion changed (range), and did the food fit the brief (appropriateness). That's what Design Creativity Bench does to frontier AI models, except the medium is UI design. The committed claim: frontier AI models can now generate functional UIs, but they cannot make creative design decisions at anything close to human level. The best model tested — OpenAI's GPT-6 Astra — scores 65.5 on the composite creativity metric against a human designer reference of 92.3, a 29% gap. This is not a coding benchmark or a pixel-fidelity test. It measures whether models can think divergently about visual and interaction design when given open-ended briefs. The benchmark structure is the real contribution. Eighty open-ended UI tasks, each adapted to two distinct product domains, scored across three dimensions. Originality measures how distinct a model's designs are from other models' outputs on the same brief — essentially, are all sixteen models converging on the same solution? Creative Range measures whether a model changes its design decisions when the product domain shifts. Appropriateness measures functional fit to the brief. The three-dimensional structure exposes something a single-metric benchmark cannot: models can be appropriate (functional) while being nearly identical to each other (low originality) and rigid across contexts (low range). Sixteen frontier models were evaluated, which gives the leaderboard real density. The paper is published alongside a reproducible evaluation framework, HuggingFace datasets, and GitHub code. The comparison baseline is human designers — not previous AI benchmarks — which is the right ladder to use for a creativity claim. The gap is large enough that it's not within noise, but the paper's honesty depends on how the human reference scores were collected and whether those designers were given the same constraints the models were. The architectural insight is that this is a behavioral benchmark, not a model-architecture paper. It doesn't propose a new model — it proposes a new measuring instrument. The instrument's value depends entirely on whether the three dimensions (originality, range, appropriateness) actually decompose creativity in a way that's stable, reproducible, and not gameable. Originality measured as inter-model distinctiveness is clever but has a vulnerability: as models diversify, a model could score high on originality by being bad in unique ways. The integrity picture is mixed in revealing ways. The evaluation framework is public, the datasets are on HuggingFace, and the code is on GitHub — all strong signals. But the benchmark is self-authored by Kombai, a company that builds AI-to-code design tools, which means the benchmark creator has commercial interest in demonstrating that AI design creativity is measurable and improvable. The human reference at 92.3 is a single number with no reported variance or sample size for the human designers. Pre-registration is absent. The real milestone question is whether this benchmark will be adopted by other labs as a standard evaluation. The 65.5-to-92.3 gap gives the field a concrete number to chase, but the benchmark's durability depends on whether originality, range, and appropriateness prove to be the right decomposition — or whether they're proxies that can be gamed. The obvious experiment not run: testing whether fine-tuning on the benchmark's own evaluation criteria closes the gap, which would tell us whether the deficit is fundamental or trainable. Kombai likely didn't run it because a positive result would undermine the benchmark's value as a long-term measuring stick, and a negative result would be their next paper.
