The AI leaderboard for 3D, ranked by votes in Spline Arena

Two models get the same prompt and each builds a scene in Spline. You vote for the better one, and every vote moves the rankings.

Benchmark

A leaderboard for 3D, voted by people

Every vote adds up to the rankings: which models build the best scenes, and which do it for less.

#ModelRatingvs. avgW / T / LBoth badCostTime
1Claude Fable 5.21621 ±11+15.4%62% won, 15% tied, 22% lost6.7%$5.6314m 43s
2Gemini 4 Carbon1589 ±12+11.6%58% won, 16% tied, 27% lost7.4%$1.6023m 37s
3Gemini 4 Argon1587 ±9+11.3%60% won, 14% tied, 27% lost6.3%$2.0116m 9s
4Gemini 4 Hydrogen1568 ±11+8.8%55% won, 17% tied, 29% lost7.3%$0.987m 1s
5Claude Opus 5.51565 ±8+8.5%56% won, 15% tied, 29% lost6.7%$2.279m 31s
6GPT-6 Astra1489 ±8−1.5%44% won, 16% tied, 40% lost7.2%$3.775m 29s
7GPT-6.1 Sol1462 ±7−5.0%40% won, 16% tied, 44% lost7.4%$0.494m 26s
8Claude Fable 5.11462 ±8−5.0%40% won, 17% tied, 43% lost9.1%$4.0710m 32s
9Claude Sonnet 5.51442 ±8−7.6%37% won, 16% tied, 47% lost8.7%$1.287m 12s
10Kimi K3.1 Pro1410 ±10−11.7%30% won, 20% tied, 50% lost12.0%$1.576m 11s
11Gemini 3.8 Flash1405 ±8−12.2%31% won, 16% tied, 53% lost9.9%$0.547m 51s
12Grok 4.71386 ±8−14.5%28% won, 17% tied, 55% lost10.7%$1.6214m 52s
13GLM 5.31332 ±11−20.9%21% won, 18% tied, 61% lost11.8%$0.6713m 39s
14DeepSeek V4 Pro1290 ±13−25.2%17% won, 18% tied, 65% lost11.5%$0.3110m 48s

Head-to-head records and score vs. cost in the full rankings.

How it works

Two models, one prompt, your vote

Look at two finished scenes, pick the better one, and every vote feeds the rankings.

  1. 01 · Prompt

    Two scenes

    Two models built a 3D scene from the same prompt using Spline.

  2. 02 · Vote

    Your pick

    Orbit both and pick the better one: A, B, a tie, or both bad.

  3. 03 · Claim

    Free AI credits

    Reach the day’s goal and claim free AI credits.

Why Spline Arena

Better models, better results

  1. See what works

    Every vote shows which models build 3D well, and where they still fall short.

  2. Improve the models

    That signal helps the models, and the scenes they make in Spline, get better.

  3. Unlock creative freedom

    Better models mean you get better results in Spline, faster.

FAQ

Questions, answered

Do I need an account to vote?

Yes, a free Spline account. It lets us monitor voting, and lets you claim free AI credits.

Which models are in the arena?

Models from Anthropic, OpenAI, Google, xAI and others. The lineup changes as new models ship, and the Rankings list every released model competing.

How are the rankings calculated?

Each score is how much more often a model wins a blind battle than the average model would: +7% means it wins 57% of the time, where an average model wins 50%, and ties and “both bad” count as half a win. A Bradley–Terry model is fitted to every head-to-head vote, and bootstrapping gives each score a 95% confidence interval. When two models’ ranges overlap a lot, treat them as tied.

How do rewards work?

Reaching the day’s vote goal lets you claim free AI credits in the arena. Take your time and consider each vote, as careful votes carry more weight.

Why is Spline running Spline Arena?

To find out which models build 3D well and help them improve, so you get better results, faster.

What happens to my votes?

They feed the public Rankings and let you claim free AI credits. They may also be shared, de-identified, with AI research partners.

Cast your first vote

Sign in with a free Spline account and help rank the models that build in 3D.