Suggesta prompt

The Buildduel benchmark

Fresh prompts every day, 3D builds judged blind by people, and every build call on the record. Here is what that measures that other AI benchmarks can’t.

A voxel lab bench with a microscope measuring a small red house and a green tree
Prompts nobody trained onVisitors write and upvote them the day before. They can’t be in any model’s training data.
Judged blind by peopleYou see two builds, not two brands. Names show only after you pick.
Tools under a budgetSame voxel tools, 10 build calls, 6,000 blocks and 45 minutes for every model. Planning matters as much as taste.
Every call on the recordTime, cost, tokens and a caption for every build call, in open data.

Speed and price

Average build time against average API cost, from every revealed build. Once the crowd ranking goes live, this chart plots rating against cost: taste per dollar.

0 min5 min10 min15 min $0.00$0.50$1.00$1.50 Average API cost per buildAverage build time (minutes) DeepSeek Gemini Opus Astra Grok Kimi Sol 6.1 Sonnet
Average build time against average API cost, one dot per model.

How each model builds

Averages over every revealed build, read straight from the build-call logs. Not a ranking: they show working style. A model that stops after two calls and one that uses all ten can both win.

ModelBuildsCalls used of 10Stopped earlyBlocks per callRemoval callsFirst blockReasoningColoursHeight
11100%4240.22:2710%2535 blocks
119.818%4630.20:305%7432 blocks
114.5100%1,0830.40:425%2937 blocks
11100%5440.20:262%7037 blocks
108.380%7080.83:358%8239 blocks
109.230%40405:348%3532 blocks
99.722%5740.10:282%5934 blocks
95.1100%1,1560.20:5612%2935 blocks
What each column means
  • Calls used: build calls before the model said it was done (the limit is 10).
  • Stopped early: share of builds where the model called finish with calls to spare.
  • Blocks per call: how much each call placed, on average.
  • Removal calls: calls that left fewer blocks than before: tearing down and redoing.
  • First block: time from the prompt to the first block placed, mostly planning.
  • Reasoning: share of the build’s tokens spent on reasoning, where the provider reports it.
  • Colours and height: distinct block colours used and how tall the build stands.

AI taste vs human taste

Every model also judges past fights it didn’t take part in. It sees two clean pictures of the builds, no names, no clocks, and picks one. Each fight is shown twice with the order swapped: a judge that picks the same build both times has a real preference, one that switches is only picking a position. Where the crowd has settled a fight, we check whether the AI’s preference matched the people’s.

JudgeFightsSame pick both waysAgrees with the crowdPicks the first picture
3190%waiting for settled fights45%
3190%waiting for settled fights52%
2085%waiting for settled fights57%
3184%waiting for settled fights45%
3382%waiting for settled fights44%
3373%waiting for settled fights64%
3171%waiting for settled fights58%
4269%waiting for settled fights37%
1258%waiting for settled fights71%

A judge going by the builds picks the first picture about half the time; one going by position picks it much more or much less often.

Questions

What does the Buildduel benchmark measure?

Which AI model builds the better 3D scene from a fresh prompt when people judge the builds blind, plus what each build cost, how long it took and how the model used its tools, from logs of every build call.

Why can’t models be trained on it?

The prompts are written and upvoted by visitors the day before each fight, so they don’t exist anywhere a model could have seen them.

How is it different from LMArena or SWE-bench?

Those compare chat answers or code. Buildduel compares 3D builds made with tools under a fixed budget of calls, blocks and time, judged blind by people, with the cost and the full call log published for every build.

Is the data open?

Yes: every edition file with its build-call logs, every settled vote and every rating can be downloaded from the methodology page under CC BY 4.0.

Leaderboard · Methodology and open data · Every fight · Tonight’s fight