Design Arena is the world's first large-scale crowdsourced benchmark platform designed specifically to evaluate the aesthetic and functional capabilities of artificial intelligence models in design-related tasks. Created by Intelligence, the platform aims to move beyond purely technical metrics by measuring how well AI models exhibit 'taste' and creative competence, as judged directly by a global community of users. By presenting the same creative prompts to multiple models simultaneously and collecting blind, side-by-side votes, the platform generates comprehensive, data-backed leaderboards that reflect real-world performance rather than cherry-picked examples.

The core functionality of the platform relies on a sophisticated tournament-style evaluation process. When a user submits a creative prompt, the system selects four models from the active pool to participate in a series of head-to-head battles. These battles occur in multiple stages—initial rounds, winners' brackets, losers' brackets, and tie-breakers—ensuring a complete ordering of the four models for every single session. Because model identities remain hidden throughout the evaluation, the resulting rankings are protected against brand bias, ensuring that the community-driven data provides an honest reflection of current AI design performance.

Some of the key features are:

• Crowdsourced Rankings: Leaderboards are dynamically updated based on millions of anonymized pairwise comparisons from users in over 190 countries.
• Tournament Evaluation: A five-battle tournament format ensures consistent and statistically meaningful ranking data for every model evaluated.
• Bradley-Terry Scoring: Employs the rigorous Bradley-Terry statistical model to calculate inherent 'strength' and ratings for each AI model.
• Blind Testing: Ensures unbiased evaluation by hiding the identities of AI models during the voting process.
• Comprehensive Metrics: Provides detailed data on model pricing, context window sizes, maximum output tokens, and specific domain win rates.
• Domain-Specific Leaderboards: Tracks performance across fourteen distinct buckets, including UI components, 3D design, website generation, and game development.

Operationally, the platform functions by taking a user-provided prompt, selecting a set of AI models, and distributing the request to them in parallel. Users view the generated outputs side-by-side and cast votes for the preferred result. These pairwise votes feed directly into the system's Elo-based ranking algorithms. This continuous cycle of crowdsourced feedback allows the platform to maintain up-to-date performance snapshots of the most advanced models available today, providing developers and designers with actionable insights into which tools excel at specific creative tasks.

Some common use cases include:

• Model Selection: Identifying the best-performing AI models for specific design domains, such as SVG generation or full-stack web application development.
• Performance Benchmarking: Comparing the cost-to-performance ratio of different AI models based on their input/output costs and overall win rates.
• Taste Evaluation: Using human judgement to determine which AI models align best with professional aesthetic standards and design requirements.
• Feature Comparison: Analyzing how different models handle complex tasks like agentic game development, data visualization, or UI component generation.