Support

Settings

Theme

Explore themes

Mode

Library

Looks borrowed from benchmark sites and products. Themes with a light and a dark version follow your mode.

Bring your own theme

Upload a Priors theme file, or have your agent make one: copy the prompt, tell it the look you want, and upload what it returns. With the Priors skill installed it already knows the format.

Custom themes stay in this browser. They travel inside shared chart links and a page's theme, so others see them as you do.

For the builder, Explore and model pages on this device. Published pages keep the theme their owner picked.

Mode

Update notices

A note when a benchmark one of your indexes uses gets a new version, with a one-click upgrade.

Benchmarks · Vision · Reasoning & puzzles

ShadowBench

Hand shadows: shown a plate from a nineteenth-century book, the model poses two 3D hands between a lamp and a screen, one tool call at a time, until their shadow matches the animal. Poses a real hand could not make are refused. Scored 0 to 100 on overlap and outline against the target, averaged over 20 plates, with token cost per episode. No model has solved a plate. From Perpetual Research. ⁣Reference scores, not models: Featureless blob 56.9, Blind search (CMA-ES) 36.2, Hands at rest 12.1.

6 models · 18 results · updated 2026-11-05 · perpetualresearch.si ↗

Leaderboard

Opus 5.5 max

Score vs subscription cost per task

Highlight models

Click a point on the chart to highlight that exact effort; click it again to remove it.

Heavy line: best score at each price. Thin lines join one model's effort levels. Cost uses each model's median subscription plan, or API price with no plan data.

More from this benchmark

6 published measures

Breakdown

darker is better within each column · click a column to sort
Model
Opus 5.5 max0.8150.28520.068.3
Sonnet 5.5 max0.8150.27820.060.0
GPT-6.1 Terra max0.7720.24520.061.4
GPT-6.1 Sol max0.7660.24220.055.5
Haiku 5.5 max0.7030.2120.060.0
GPT-6 Luna max0.6090.14820.049.3

All results

18 rows
#ModelEffortShadowBenchSubscription $ / taskAPI $ / taskAgent stepsTime
1Opus 5.5max48.8 ±2.8$0.11$7.139754 min
2Sonnet 5.5max48.8 ±2.3$0.079$5.239760 min
3GPT-6.1 Terramax46.4 ±2.7$0.017$0.449922 min
4GPT-6.1 Solmax44.7 ±2.8$0.030$0.8810040 min
5GPT-6.1 Terrahigh44.4 ±3.1$0.008$0.219823 min
6Opus 5.5high43.6 ±2.9$0.019$1.279323 min
7GPT-6.1 Terramedium43.4 ±1.8$0.007$0.189725 min
8GPT-6.1 Solhigh41.9 ±2.7$0.013$0.3910035 min
9GPT-6.1 Solmedium41.4 ±3.1$0.011$0.329939 min
10Haiku 5.5max41.3 ±3.7$0.007$0.349554 min
11Sonnet 5.5high40.6 ±2.6$0.006$0.416513 min
12Opus 5.5medium39.8 ±3.5$0.013$0.858522 min
13Haiku 5.5high35.6 ±3.1$0.001$0.0366412 min
14GPT-6 Lunamax35.1 ±3.4$0.007$0.0988924 min
15GPT-6 Lunahigh35.0 ±2.3$0.004$0.0599020 min
16GPT-6 Lunamedium31.6 ±3.0$0.003$0.0419021 min
17Haiku 5.5medium28.5 ±3.2$0.000$0.012357.4 min
18Sonnet 5.5medium17.3 ±5.3$0.000$0.030854s