Skip to content
GitHub
Sound Theme

Leaderboard

Benchmarks

How well each model draws pixel art through the plugin, on the same fixed tasks.

Leaderboard

Score is 0–100: 35% criteria passed, 50% craft from human judges, 15% speed relative to the fastest run of each benchmark. A run with no recorded time is scored on criteria and craft alone. Each model's best run per benchmark counts, averaged over every benchmark; unrated craft and benchmarks it has not run count as 0. Read the rubric.

Models ranked by their mean score out of 100
# Model Score Compliance Craft Speed Coverage Runs
1 claude-opus-5-5
96% 50% 68% 4/4 benchmarks 24
2 claude-sonnet-5-5
79% 33% 76% 4/4 benchmarks 4
3 gpt-6.1-sol
52% 20% 46% 3/4 benchmarks 3
4 gpt-6-luna
43% 3% 68% 4/4 benchmarks 4

Score over time

Each model's leaderboard score from the first run to today (a year at most). A marker is a day the model ran; the line is solid between runs and dashed where it did not run, carrying its nearest score. Benchmarks it had not run yet count 0, so a line climbs both when the model draws better and when it covers more of the suite.

Leaderboard score out of 100 for each model, by day
SeriesOct 1, 2026Oct 2, 2026Oct 5, 2026Oct 6, 2026Oct 7, 2026
claude-opus-5-54144636869
claude-sonnet-5-5——55——
gpt-6.1-sol——35——
gpt-6-luna——27——

All benchmarks

  1. 1

    Fixed prompts

    Same words, canvas and palette for every model.

  2. 2

    Run in Aseprite

    The model draws through the plugin, one session per step.

  3. 3

    Scored

    Each criterion is a yes-or-no check on the finished files.

Benchmarks are run by the maintainers; new models are added as they are tested. A prompt's revision changes when its wording or criteria do, and only runs on the current revision are ranked.