OpenArt Launches Creative AI Arena That Splits Model Rankings by Job

OpenArt’s new Arena promises a more useful answer than a single “best model” score: choose an AI image or video model for the creative job at hand. Its launch rankings do show different winners across tasks, but the benchmark will keep some active prompts private—putting transparency at the center of whether creative teams treat it as a decision tool.

OpenArt Arena is a public benchmark for AI image and video generation. Rather than collapse performance into one ranking, it publishes separate leaderboards for filmmaking, advertising, graphic design, e-commerce, motion design, video editing, lip sync and overall performance.

A benchmark built around the production task

The premise is straightforward: a model that handles cinematic lighting and camera movement well may not be the best fit for an ad that needs readable text, accurate logos and faithful product placement. OpenArt generates outputs from curated prompts for each board, then presents unlabeled pairs to evaluators for side-by-side choices.

Those preferences are aggregated with the Bradley-Terry model, a statistical method that estimates an entrant’s chance of winning from head-to-head comparisons. OpenArt planned to draw on a Creative Expert Council and roughly 800 to 1,000 additional “tastemakers,” though it did not disclose how many people completed the launch voting.

1,081Overall video leader

Seedance 2.5 ranked first on the overall video board, ahead of Wan 3.0 at 1,004 and Seedance 2.0 at 1,000.

1,034Video editing result

Wan 3.0 placed first for video editing by one point over Seedance 2.5, a gap OpenArt’s confidence intervals caution against treating as a meaningful quality difference by itself.

1,010Overall image leader

Seedream 5.0 Pro led the overall image ranking, 10 points ahead of GPT Image 2.

The launch results make the case—and limit it

The video table produces a strong general leader. Seedance 2.5 topped the overall board and ranked first on four of the five specialized video leaderboards supplied for launch. But Wan 3.0 edged it in video editing, illustrating why an overall winner may not settle a team’s choice for a specific workflow.

The image boards are more divided. GPT Image 2 ranked first for graphic design and image editing, while Seedream 5.0 Pro led the film-oriented imagery and e-commerce boards, as well as the overall image ranking. That pattern is the practical value of task-specific comparison: it can direct users toward candidates rather than declare one universal champion.

The unanswered question is what the judges saw

OpenArt says it will publish its methodology and part of its prompt sets, while withholding some active prompts to make it harder for model makers to tune directly for the test. That protects against a static benchmark becoming a target. It also means users cannot fully inspect whether the hidden prompts represent the kinds of creative work they need to do.

For now, Arena is most useful as a structured starting point, not a final procurement verdict. The close video-editing result also shows why teams should read the task board and its confidence intervals, rather than treat small score differences as proof that one model is materially better.