OpenArt launches task-specific benchmark for AI image and video models

OpenArt, an artificial intelligence startup founded by former Google employees, has launched a new public benchmark called OpenArt Arena. [1] The platform provides specialized leaderboards to help creatives select the best image and video generation models for their specific needs. [2] Instead of assigning a single overall score, the benchmark evaluates models across distinct categories including filmmaking, e-commerce, graphic design, and video editing. [1][2]

The arena relies on blind pairwise evaluations conducted by creative practitioners to rank the competing generation tools. [1] In the initial rankings, ByteDance's Seedance 2.5 secured the top position in the video category. [2] Meanwhile, Alibaba's Seedream 5.0 Pro leads the leaderboard for image generation. [2] To prevent developers from over-optimizing their systems for the test, OpenArt is keeping certain evaluation prompts private while publishing its overall methodology. [1][2]

The introduction of task-specific leaderboards reflects a maturing generative media market where overall capability matters less than specialized workflow performance. By fragmenting the benchmark into specific creative jobs like lip sync and motion design, OpenArt Arena directly challenges the utility of broad, single-score evaluations for multimodal models. The early dominance of ByteDance and Alibaba models highlights fierce competition for established tools like OpenAI's GPT Image series and Google's Nano Banana family. What the sources don't address: how frequently the leaderboards will be updated as new proprietary and open-source models are released.