Design Arena launches as crowdsourced AI design benchmark

Design Arena, a startup from Y Combinator’s Summer 2025 batch, launched a public benchmark for AI-generated visual and interface design. The site gives the same creative prompt to several AI models, shows the resulting outputs side by side to a visitor, and asks them to vote for the one they judge better — a head-to-head format modelled on the “arena” approach earlier used for text-based chatbot comparisons. Votes were aggregated using Bradley-Terry-style scoring, similar to an Elo rating, across categories including websites, user interfaces, games, 3D output and other generated media.

The benchmark addressed a gap other evaluation methods had struggled with: aesthetic and design quality resist the kind of objective scoring used for math or coding benchmarks, since there is often no single correct answer, only a comparative judgement of taste. By putting that judgement to a large number of anonymous human voters rather than a fixed rubric, Design Arena offered labs a signal for how their models performed on subjective, presentation-facing tasks that other benchmarks did not measure.

The site grew quickly after launch, reportedly reaching several million users across many countries within its first year, and its rankings were cited by outside observers as one of the few available comparative signals for design-oriented model output, alongside similar attention from frontier labs including OpenAI and Anthropic. The company went on to raise outside funding to expand the benchmark.