Nooks

Physical intelligence leaderboard.

Compare how AI systems design brick models, follow assembly steps and create functional objects.

How rankings work

Design intelligence

BrickBench · 12 systems · 300 design prompts

Highest design ratingGPT-6 Astra1,299 Elo · $4.06 per build
Best under $1GPT-6.1 Sol1,286 Elo · $0.87 per build
Lowest build costGLM 5.3 Flash753 Elo · $0.54 per build

BrickBench / Overall

Design quality vs. build cost

One dot per system. Higher scores mean better designs. Lower costs sit farther left.

Design rating (Elo)

80010001200$0.5$1$2$5$10$20GPT-6 Astra · 1299 Elo · USD 4.06Claude Opus 5.5 · 1287 Elo · USD 6.33GPT-6.1 Sol · 1286 Elo · USD 0.87Claude Opus 5 · 1108 Elo · USD 16.47BrickWright (Astra) · 1107 Elo · USD 1.43Qwen 3.8 Flash · 1057 Elo · USD 1.75Gemini 3.8 Flash · 1024 Elo · USD 3.70DeepSeek V4.1 Flash · 1015 Elo · USD 0.71GPT-5.6 Sol · 988 Elo · USD 1.02GPT-5.6 Luna · 879 Elo · USD 0.75Muse Spark 1.3 · 878 Elo · USD 6.68GLM 5.3 Flash · 753 Elo · USD 0.54

Average token cost per build (USD, log scale)

Best rating within budget

1,299 Elo · $4.06 per build

System rankings

SystemElo ± 95% CI$/build
GPT-6 Astra1299 ± 274.06
Claude Opus 5.51287 ± 276.33
GPT-6.1 Sol1286 ± 260.87
Claude Opus 51108 ± 2016.47
BrickWright (Astra)1107 ± 181.43
Qwen 3.8 Flash1057 ± 201.75
Gemini 3.8 Flash1024 ± 203.70
DeepSeek V4.1 Flash1015 ± 200.71
GPT-5.6 Sol988 ± 191.02
GPT-5.6 Luna879 ± 200.75
Muse Spark 1.3878 ± 236.68
GLM 5.3 Flash753 ± 290.54

Assembly rankings

BC-Bench · Five models following the same brick assembly instructions.

Overall averages from Table 2 of BC-Bench
RankModel Selection % Position error LDU Rotation error ° Step success %
1GPT-5.443.88310.7874.850.18
2Qwen-3.5-VL-27B37.44314.3282.940.18
3Qwen-3-VL-8B22.76210.1462.470.36
4InternVL-3.5-8B13.24221.6988.430.00
5Gemma-3-12B4.35269.0963.000.09

Liu et al., BC-Bench (2026), Table 2. Scores are the authors’ overall averages across objects.

Functional design

LMBuild evaluates physical products across soundness, affordance, design and realization.

LMBuild / Task results

Bar stool operability

Tier B · Name and image input · Final checkpoint

  1. Claude Fable 5.10.666667
  2. Claude Haiku 4.50.666667
  3. Claude Opus 50.666661
  4. Claude Sonnet 50.5
  5. InternVL 3.5 8B0.5

Five highest reported R.3 operability scores. The evaluator marked all five as degraded.

How to read the rankings

Each discipline uses the measures published by its research teams. Compare systems within the same benchmark and evaluation setting.

Design quality and cost

BrickBench assigns Design Elo from pairwise judgments of build quality. A higher rating means the judges more often preferred the system’s designs. The table gives each estimate and its 95% interval. Nearby point estimates can have overlapping intervals.

Each step marks the best published Design Elo available at that mean cost per build. The line rises when a more expensive system sets a higher score. Horizontal spacing represents proportional changes in mean token cost per assembly.

Assembly and function

BC-Bench reports average performance on assembly steps. Selection and success improve as scores rise. Position and rotation errors improve as they fall.

The LMBuild view shows operability on one bar-stool task under a fixed input condition. The studies below document the evaluation conditions.

Research sources (5)
BrickBench
Design · Prompt-based design, pairwise ratings and cost.
BC-Bench
Assembly · Brick selection and placement. Table 2 supplies the assembly rankings.
LTRON
Assembly · Research context for interactive reconstruction.
LEGO-ECA
Assembly · Research context for detecting and correcting assembly errors.
LMBuild
Function · Physical product evaluation. The function view uses the bar stool task’s R.3 metric.