Frontier3DBench

Benchmarking Coding Agents for 3D Creation
across Objects, Scenes, and Interactive Open Worlds

145 tasks · 9 categories · 3 levels

Leaderboard

Score / 100 · Higher is better

Total = 30% L1 + 30% L2 + 40% L3 · Scores / 100

1GPT-6.1 Sol
2GPT-6 Astra
3Claude Opus 5.5
4DeepSeek V4.1 Flash
5Kimi K3.1 Pro
6MiniMax M3.1 Flash
7Kimi K3
8GPT-6 Luna
9GLM-5.3-Flash
10Qwen 3.8 Max
11Doubao Seed Evolving
Full results L1 · L2 · L3 · Total
145-task results. L1: 65 tasks, L2: 65 tasks, L3: 15 tasks. Bold scores lead each column.
#ModelEffortL1 65L2 65L3 15Total ↑Harness
1GPT-6.1 Solmax 71.2170.3558.8966.02 Codex
2GPT-6 Astramax 68.6362.5161.2863.85 Codex
3Claude Opus 5.5max 67.3762.6158.2362.29 Claude Code
4DeepSeek V4.1 Flashmax 63.2062.0049.7257.45 DSH
5Kimi K3.1 Promax 61.3459.4249.3455.96 Kimi CLI
6MiniMax M3.1 Flashmax 58.9860.5541.7452.55 Claude Code
7Kimi K3max 52.4555.5341.9649.18 Kimi CLI
8GPT-6 Lunamax 56.4054.1438.5048.56 Codex
9GLM-5.3-Flashmax 56.0155.9235.0847.61 ZCode
10Qwen 3.8 Maxxhigh 50.1950.6439.4446.03 Claude Code
11Doubao Seed Evolvingmax 51.2449.5831.7242.94 Claude Code

The benchmark

From a single object to a world that responds.

Frontier3DBench tests coding agents on the full stack of 3D creation: detailed objects, coherent scenes, and interactive worlds. Agents build native Blender and Unreal Engine projects that can be opened, edited, and executed.

145tasks
3levels
9task families
L1

Objects

Geometry · Structure · Appearance

65 tasks 3 families
L2

Scenes

Layout · Appearance · Editing

65 tasks 4 families
L3

Interactive open worlds

Connected regions · Object interactions · Persistent state

15 tasks 2 families

Task coverage

LevelTask familyInputTasks
L1Text2CodeText specification15
L1Img2CodeMulti-view images + prompt25
L1Img2Code-SingleViewSingle image + prompt25
L2Brief2SceneText brief20
L2Ref2SceneRendered image + prompt10
L2Ref2Scene-RealPhotoReal photograph + prompt25
L2SceneEditInitial scene + reference + edit10
L3Text2EnvText brief5
L3Img2EnvImages + brief10
Total145

From reference to executable 3D

L1Reference to object

INPUT
Original typewriter task reference
Reference image
Coding agent
OUTPUT Editable
◉Blender (L1)Typewriter.blend−   □
FileEditLayoutModelingShadingObject Mode
Generated Typewriter inside an illustrative Blender viewport
User Perspective
◂ ▷ ▸
020406080100120
Meshes · Materials · Modifiers.blend

L2Reference to scene

INPUT
Original museum hall task reference
Reference image
Coding agent
OUTPUT Editable
◉Blender (L2)Museum.blend−   □
FileEditLayoutModelingShadingObject Mode
Generated Museum Hall inside an illustrative Blender viewport
User Perspective
◂ ▷ ▸
020406080100120
Objects · Lighting · Materials.blend

L3Reference to environment

INPUT
Original interactive workshop task reference
Reference image
Coding agent
OUTPUT Editable
UUnreal Engine 5 (L3)Workshop.uproject−   □
FileEditSelectLandscapeBlueprints▶ Play
Generated BP_Workbench inside an illustrative Unreal Engine 5 viewport
Perspective · Lit
Content Browser
▰Maps▰Meshes▰Materials▰Blueprints
Levels · Blueprints · Interactions.uproject

Built to run. Measured in 3D.

Evaluation metrics ↗

Native projects connect visual quality to real geometry and executable behavior. Independent evaluation measures what agents build, how it looks, and how it works.

Visual fidelityStructural consistencyTask fulfillment

Inside the results

66.02

Best overall

Sol leads the complete benchmark.

61.28

Interactive worlds

Astra leads both L3 task families.

84.05

Text-to-object

Opus leads specification-driven modeling.

Quality and cost efficiency

Luna, DeepSeek, and Sol trace the quality–cost frontier at the paper’s estimated API costs.

Paper figure 5: model score versus estimated API cost per task, with Luna, DeepSeek, and Sol on the Pareto frontier.

Appearance. Structure. Task fulfillment.

Three capability profiles distinguish visual fidelity, structural consistency, and completion of the requested task.

Paper figure 4: visual quality, structural consistency, and task fulfillment by model across L1, L2, and L3.
Explore all nine task families Model profiles ↗

Strengths across nine task families

Paper figure 3: eleven model radar profiles across nine task families, compared with the cross-model average.

Gallery

L1ObjectsDrag to rotate · Scroll to zoom

Typewriter

GPT-6 Astra
Typewriter generated by GPT-6 Astra 76.69/ 100
GPT-6.1 Sol
Typewriter generated by GPT-6.1 Sol 70.98/ 100
Claude Opus 5.5
Typewriter generated by Claude Opus 5.5 73.53/ 100

Microscope

GPT-6 Astra
Microscope generated by GPT-6 Astra 70.28/ 100
GPT-6.1 Sol
Microscope generated by GPT-6.1 Sol 78.55/ 100
Claude Opus 5.5
Microscope generated by Claude Opus 5.5 66.00/ 100

Sneaker

GPT-6 Astra
Sneaker generated by GPT-6 Astra 24.82/ 100
GPT-6.1 Sol
Sneaker generated by GPT-6.1 Sol 74.96/ 100
Claude Opus 5.5
Sneaker generated by Claude Opus 5.5 52.74/ 100

L2Scenes

Museum hall

GPT-6 Astra
62.16/ 100
GPT-6.1 Sol
57.40/ 100
Claude Opus 5.5
49.96/ 100

Antique shop

GPT-6 Astra
28.66/ 100
GPT-6.1 Sol
64.49/ 100
Claude Opus 5.5
61.01/ 100

Classroom

GPT-6 Astra
61.56/ 100
GPT-6.1 Sol
65.76/ 100
Claude Opus 5.5
50.35/ 100

L3Interactive worlds

Frontier town

GPT-6 Astra
56.04/ 100
GPT-6.1 Sol
52.18/ 100
Claude Opus 5.5
54.16/ 100

Minecraft village

GPT-6 Astra
71.89/ 100
GPT-6.1 Sol
70.29/ 100
Claude Opus 5.5
76.44/ 100

Workshop

GPT-6 Astra
63.72/ 100
GPT-6.1 Sol
56.82/ 100
Claude Opus 5.5
58.67/ 100