SplineIntroducing Spline Arena
The AI leaderboard for 3D, ranked by your votes. Two models build a scene in Spline from the same prompt, you pick the better one, and every vote moves the rankings.
Today we're opening Spline Arena to everyone. It's a public leaderboard for how well AI models build in 3D, and the rankings come from you.
AI models are getting good at code and text, and benchmarks for those are everywhere. 3D is different. Whether a scene is good depends on how it looks, how it's lit, how its parts fit together and whether it does what the prompt asked. That's hard to score automatically, and easy for a person to judge at a glance. So we built an arena where people do the judging.
How it works
Two models get the same prompt, like "A bicycle you could ride to work", and each builds a real scene in Spline.
You see both finished scenes side by side. Orbit them, pan around, look at them from every side, then pick the better one: A, B, a tie, or both bad. The models stay anonymous, so a name never sways the pick, and the next battle loads as soon as you vote. How the models compare shows in the Rankings.
The prompts range from a single object to whole interactive scenes, with lighting, animation and things to click:
The rankings
Every vote feeds the public Rankings. We fit a Bradley–Terry model to every head-to-head vote and bootstrap it to give each score a confidence interval. When two models are close, the rankings show it.
Each score reads as how much more often a model wins a blind battle than the average model would. A score of 7% means it wins 57% of its battles where an average model wins 50%. Ties and "both bad" count as half a win for each side.
The arena compares frontier models from Anthropic, OpenAI, Google, xAI and others, and the lineup changes as new models ship. Next to the leaderboard, a chart plots each model's score against its average cost per scene, so you can see which ones give the most for the money.
Why we're doing it
Spline is where more and more people build 3D with AI, in our own agent and through Claude, ChatGPT, Cursor and other MCP clients. We want to know which models build 3D well and where they still fall short, and we want that answer to come from the people who use them.
That signal helps the models get better at 3D, and the scenes they make in Spline get better with them. Votes may be shared, de-identified, with AI research partners for exactly that reason. Better models mean better results for you, faster.
Get started
Spline Arena is in the dashboard sidebar, in the new tab launcher, and in its own tab in the desktop app. All you need is a free Spline account. Learn more at spline.design/arena.
