ShadowBench
Models pose two 3D hands between a lamp and a screen, one tool call at a time, until their shadow matches one of 20 hand-shadow plates from books of 1859 to about 1900. The environment refuses any pose real hands can't make. Score, out of 100, mixes the overlap with the target and how closely the outlines match.
Each model ran once per plate and effort, so small differences are not resolved. No model solved a plate. A featureless blob, the convex hull of each target, scores 56.9, above every model's average, and blind search over the pose scores 36.2. Models rarely hold the hands together, as the books do. Six models from two labs were tested, with no human baseline yet.
- Made by
- Perpetual Research
- Released
- November 2026
- Kind
- Spatial reasoning and counting
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Strongest model tested
- Claude Opus 5.5 #4 on the capability index
- Last model added
- October 2026
- Scores refreshed
- November 12th, 2026
Claude Opus 5.5 leads with 48.8
Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
- 1 Claude Opus 5.5 max effort 48.8
- 2 Claude Sonnet 5.5 max effort 48.8
- 3 GPT-6.1 Terra max effort 46.4
- 4 GPT-6.1 Sol max effort 44.7
- 5 Claude Haiku 5.5 max effort 41.3
- 6 GPT-6 Luna max effort 35.1
Source: Official ShadowBench leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 48.8 | max effort | Sep 2026 | Official leaderboard |
| 2 | Claude Sonnet 5.5 | 48.8 | max effort | Sep 2026 | Official leaderboard |
| 3 | GPT-6.1 Terra | 46.4 | max effort | Oct 2026 | Official leaderboard |
| 4 | GPT-6.1 Sol | 44.7 | max effort | Sep 2026 | Official leaderboard |
| 5 | Claude Haiku 5.5 | 41.3 | max effort | Oct 2026 | Official leaderboard |
| 6 | GPT-6 Luna | 35.1 | max effort | Sep 2026 | Official leaderboard |
Mean score against the cost of one episode
Official score and cost plot across medium, high, and maximum reasoning effort. The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.
| Model | Setting | Mean score | Cost per episode |
|---|---|---|---|
| Claude Opus 5.5 | max effort | 48.82 | $7.13 |
| Claude Sonnet 5.5 | max effort | 48.79 | $5.23 |
| GPT-6.1 Terra | max effort | 46.38 | $0.44 |
| GPT-6.1 Sol | max effort | 44.72 | $0.88 |
| GPT-6.1 Terra | high effort | 44.38 | $0.21 |
| Claude Opus 5.5 | high effort | 43.63 | $1.27 |
| GPT-6.1 Terra | medium effort | 43.39 | $0.18 |
| GPT-6.1 Sol | high effort | 41.9 | $0.39 |
| GPT-6.1 Sol | medium effort | 41.39 | $0.32 |
| Claude Haiku 5.5 | max effort | 41.31 | $0.34 |
| Claude Sonnet 5.5 | high effort | 40.58 | $0.41 |
| Claude Opus 5.5 | medium effort | 39.83 | $0.85 |
| Claude Haiku 5.5 | high effort | 35.61 | $0.04 |
| GPT-6 Luna | max effort | 35.09 | $0.1 |
| GPT-6 Luna | high effort | 35 | $0.06 |
| GPT-6 Luna | medium effort | 31.56 | $0.04 |
| Claude Haiku 5.5 | medium effort | 28.51 | $0.01 |
| Claude Sonnet 5.5 | medium effort | 17.29 | $0.03 |
The best score, 48.8, costs $7.13 a task
Runs on the cost frontier score higher than any run that's cheaper. GPT-6.1 Terra at max effort comes within 2.4 of the best for $0.44 a task, 16 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
Source: Official ShadowBench leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| Claude Opus 5.5 | max effort | 48.8 | $7.13 | ||
| Claude Sonnet 5.5 | max effort | 48.8 | $5.23 | ||
| GPT-6.1 Terra | max effort | 46.4 | $0.44 | ||
| GPT-6.1 Sol | max effort | 44.7 | $0.88 | ||
| GPT-6.1 Terra | high effort | 44.4 | $0.21 | ||
| Claude Opus 5.5 | high effort | 43.6 | $1.27 | ||
| GPT-6.1 Terra | medium effort | 43.4 | $0.18 | ||
| GPT-6.1 Sol | high effort | 41.9 | $0.39 | ||
| GPT-6.1 Sol | medium effort | 41.4 | $0.32 | ||
| Claude Haiku 5.5 | max effort | 41.3 | $0.34 | ||
| Claude Sonnet 5.5 | high effort | 40.6 | $0.41 | ||
| Claude Opus 5.5 | medium effort | 39.8 | $0.85 | ||
| Claude Haiku 5.5 | high effort | 35.6 | $0.04 | ||
| GPT-6 Luna | max effort | 35.1 | $0.10 | ||
| GPT-6 Luna | high effort | 35.0 | $0.06 | ||
| GPT-6 Luna | medium effort | 31.6 | $0.04 | ||
| Claude Haiku 5.5 | medium effort | 28.5 | $0.01 | ||
| Claude Sonnet 5.5 | medium effort | 17.3 | $0.03 |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official ShadowBench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official ShadowBench leaderboard
No lab model card reports it yet
We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.
More in Vision
- MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy. GPT-5.1 85.4%
- MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
- CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
- MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy.
Qwen 3.8 Max 97.7% - MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy. Qwen 3.8 Max 85.9%
- ChartQA In use Models answer human-written and generated questions about chart images that need visual and arithmetic reasoning. Scored by relaxed accuracy with a 5 percent numeric tolerance.
Command A Vision 90.9%