What it measures, and what it doesn’t

How a fight runs

1. The three most upvoted prompts at 12:00 make the card. Each one is checked again before it reaches the models.
2. Two different models are drawn from the active roster for each fight, favouring models that have fought less. Which one is A and which is B is random.
3. Both get the same instructions and the same voxel tools through OpenRouter: at most 10 build calls, 6,000 blocks and 45 minutes. No temperature or other sampling setting is changed, so each provider’s defaults apply. The harness records every call, the clock and the API bill.
4. If a model fails twice in a row, another model takes its place, and the fight’s file records the substitution.
5. At 18:00 the builds go live as 3D replays. Names, time, cost and block counts stay hidden until each person picks, so the pick is about the build itself.

From picks to a rating

Each settled fight is one comparison between two models, and its result is the share of blind picks each side won, so every fight counts the same however many people saw it. A Bradley–Terry model finds the strength for every model that best explains all the fights, on an Elo-like scale where 1000 is average. A gap in rating means a predictable chance of winning:

rating(model) = 1000 + 400 · log10(p_model)        # Bradley–Terry strength p, geometric mean = 1
P(A beats B)  = 1 / (1 + 10^((rating_B − rating_A) / 400))
wins_i = Σ over i's fights of its share of the picks; n_ij = fights between i and j
fit:  p_i ← wins_i / Σ_j  n_ij / (p_i + p_j)        # 200 iterations, one pseudo-fight per pair
95% CI: 200 bootstrap resamples of fights, 2.5th–97.5th percentile

• Blind picks only. Once the next edition goes live, the archive names the models, so a pick counts only if it was made before that. Old fights stay open to pick for fun.
• Settling. A fight counts once it has at least 10 blind picks and isn’t tied. The ranking switches from the sample season to real picks once 9 fights have settled (0 so far).
• Uncertainty. Every rating comes with a 95% interval from resampling fights. Overlapping intervals mean the crowd can’t yet separate two models.
• Small samples. One pseudo-fight per pair keeps ratings finite when a model has only won or only lost. A model with fewer than 3 settled fights is marked provisional and kept off the podium.

Keeping it fair

The exact instructions

Every model gets this system message, with the day’s prompt filled in, and the same 3 tools (build, look, finish). Each turn allows up to 32,000 output tokens.

You are competing in Buildduel, a blind build battle between AI models.
Build this prompt in a 3D voxel world: "<the prompt>"

World: integer grid, x and z from -24 to 24, y from 0 to 48 (y is up, y=0 is the ground layer). Viewers look from the front (+z side) at an angle, and can orbit.
Limits, identical for every model: at most 10 build calls, 6000 blocks in total, 45 minutes of wall-clock time. Blocks over the cap are not placed.
Tools: build (place or remove blocks with block, box, line, sphere, cylinder, remove ops), look (see your build as text), finish (end your turn).
Work in stages. Give each build call a short caption; viewers read them while the build replays.
People judge blind which build is better and answers the prompt best. Your time, tokens and cost are recorded and revealed after they pick.
Call finish when you are done.

Open data

0 settled fights and 0 picks so far, updated live. Free to use with credit (CC BY 4.0).

votes.csv columns

How to cite

Buildduel (2026). Blind AI build battle votes. https://buildduel.com/methodology

See also the rules, the leaderboard and the archive.