AA-3D

Artificial Analysis leaderboard of models building interactive 3D scenes: each model turns a brief into a three.js page, then a panel votes blind between two scenes.

View on Artificial Analysis
Models
21
Best Elo
1,194
Claude Sonnet 5.5 (Max, Default Fallback)
Best open-weights
982
Kimi K3 (Max)
Cheapest of the top 5
$3.46
GPT-6 Sol (Max) · 1,123

Elo by cost, tokens or turns

Average values per scene, computed as on Artificial Analysis. The horizontal axis is reversed: towards the top right, a model reaches a higher Elo for less money, fewer tokens or fewer turns.

AA-3D · Elomost efficient ↗Average cost per scene

Models shown20 of 20

Select a model to show or hide it; hovering it brings it forward on the chart. The axes fit the models shown.

21 / 21 models

  1. #1
    Claude Sonnet 5.5Max, Default Fallback

    Anthropic

    1,19495% interval: 1,160 to 1,228

    Released: Sep 28, 2026Cost / scene: $14.63Tokens / scene: 549KMedian turns: 160Reasoning

  2. #2
    Claude Opus 5.5Max, Default Fallback

    Anthropic

    1,19095% interval: 1,155 to 1,225

    Released: Sep 22, 2026Cost / scene: $20.50Tokens / scene: 393KMedian turns: 137Reasoning

  3. #3
    GPT-6 AstraMax

    OpenAI

    1,16595% interval: 1,131 to 1,199

    Released: Sep 3, 2026Cost / scene: $18.11Tokens / scene: 153KMedian turns: 60Reasoning

  4. #4
    Claude Fable 5.1Max, Default Fallback

    Anthropic

    1,13195% interval: 1,098 to 1,165

    Released: Sep 1, 2026Cost / scene: $31.50Tokens / scene: 293KMedian turns: 138.5Reasoning

  5. #5
    GPT-6 SolMax

    OpenAI

    1,12395% interval: 1,089 to 1,156

    Released: Sep 22, 2026Cost / scene: $3.46Tokens / scene: 109KMedian turns: 73Reasoning

  6. #6
    GPT-6.1 SolMax

    OpenAI

    1,11495% interval: 1,081 to 1,148

    Released: Sep 29, 2026Cost / scene: $3.51Tokens / scene: 195KMedian turns: 56Reasoning

  7. #7
    1,03595% interval: 1,003 to 1,067

    Released: Sep 2, 2026Cost / scene: $3.21Tokens / scene: 152KMedian turns: 82Reasoning

  8. #8
    GPT-6 LunaMax

    OpenAI

    1,000Reference

    Released: Sep 22, 2026Cost / scene: $0.201Tokens / scene: 144KMedian turns: 67Reasoning

  9. #9
    Claude Haiku 5.5Max

    Anthropic

    99495% interval: 962 to 1,026

    Released: Oct 7, 2026Cost / scene: $0.489Tokens / scene: 423KMedian turns: 57Reasoning

  10. #10
    Grok 4.7Xhigh

    SpaceXSI

    99495% interval: 962 to 1,026

    Released: Sep 21, 2026Cost / scene: $14.26Tokens / scene: 232KMedian turns: 81Reasoning

  11. #11
    Kimi K3Max

    Moonshot AI

    98295% interval: 949 to 1,015

    Released: Jul 16, 2026Cost / scene: $13.62Tokens / scene: 214KMedian turns: 137ReasoningOpen weights

  12. #12
    97795% interval: 945 to 1,009

    Released: Sep 21, 2026Cost / scene: $0.444Tokens / scene: 221KMedian turns: 117ReasoningOpen weights

  13. #13
    96595% interval: 932 to 997

    Released: Sep 2, 2026Cost / scene: $2.97Tokens / scene: 183KMedian turns: 109Reasoning

  14. #14
    94895% interval: 916 to 981

    Released: Sep 2, 2026Cost / scene: $8.27Tokens / scene: 193KMedian turns: 116.5Reasoning

  15. #15
    93195% interval: 899 to 963

    Released: Aug 26, 2026Cost / scene: $1.40Tokens / scene: 243KMedian turns: 182ReasoningOpen weights

  16. #16
    86895% interval: 835 to 900

    Released: Oct 6, 2026Cost / scene: $6.15Tokens / scene: 374KMedian turns: 122Reasoning

  17. #17
    84895% interval: 814 to 881

    Released: Sep 18, 2026Cost / scene: $5.91Tokens / scene: 298KMedian turns: 169Reasoning

  18. #18
    GLM-5.3Max

    Z.AI

    78295% interval: 747 to 816

    Released: Aug 18, 2026Cost / scene: $8.28Tokens / scene: 193KMedian turns: 250ReasoningOpen weights

  19. #19
    MiniMax-M3

    MiniMax

    73995% interval: 704 to 773

    Released: Jun 1, 2026Cost / scene: $1.79Tokens / scene: 132KMedian turns: 228ReasoningOpen weights

  20. #20
    InklingXhigh

    Thinking Machines

    57995% interval: 538 to 620

    Released: Jul 15, 2026Cost / scene: —Tokens / scene: 44.5KMedian turns: 65ReasoningOpen weights

  21. #21
    56995% interval: 524 to 614

    Released: Jun 4, 2026Cost / scene: $4.55Tokens / scene: 147KMedian turns: 79.5ReasoningOpen weights

Win rate by domain

Share of panel votes each model won in each of the seven task domains. A hatched cell rests on too few scenes or opponents to be reliable. Share of votes won in the domain. Based on 10,000+ panel votes.

Win rate by domain
ModelLandmarks and real placesMachinesNatural phenomenaLiving bodiesGames, sports and pilotingSpaceStories and fantasy
Claude Sonnet 5.5 (Max, Default Fallback)77.9%77.6%76.0%78.5%73.0%72.7%77.9% (Sample too small, Too few scenes (4))
Claude Opus 5.5 (Max, Default Fallback)71.4%79.9%78.2%70.5%86.0%73.3%78.8%
GPT-6 Astra (Max)82.2%76.3%63.2%77.0%66.0%80.4%76.9%
Claude Fable 5.1 (Max, Default Fallback)72.0%68.3%64.4%73.6%75.3%64.3%72.3%
GPT-6 Sol (Max)70.4%73.1%77.7%65.2%59.8%63.0%75.5%
GPT-6.1 Sol (Max)72.1%66.9%61.7%78.6%63.5%68.8%77.9%
Muse Spark 1.3 (Max)59.1%58.0%61.5%64.9%54.1%56.0%63.6%
GPT-6 Luna (Max)60.0%52.8%54.7%48.4%46.9%53.2%60.4%
Claude Haiku 5.5 (Max)47.5%48.1%51.1%58.2%60.7%55.9%54.2%
Grok 4.7 (Xhigh)55.1%45.8%50.3%55.3%59.3%56.9%54.2%
Kimi K3 (Max)52.9%55.8%51.3%50.0%50.4%43.1%59.2%
MiMo-V2.6-Pro53.0%48.0%55.2%45.1%62.2%56.0% (Sample too small, Too few scenes (4))42.7%
Gemini 3.8 Flash (High)43.5%43.2%60.8%51.3%47.5%46.7%51.0%
Qwen3.8 Max (0902)47.0%29.6% (Sample too small, Too few scenes (4))54.5%48.0%60.3%45.5%36.3% (Sample too small, Too few scenes (4))
GLM-5.3-Flash39.0%41.2%49.0%40.9%51.9%46.6%46.9%
Mistral Large 4 Preview31.1%29.5%38.5%46.2%35.6%50.0%38.8%
Step 5 Preview39.0%35.4%31.8%39.0%36.2%32.7%34.5%
GLM-5.3 (Max)31.6%32.6%22.1%29.7%29.0%38.3% (Sample too small, Too few scenes (4))15.4% (Sample too small, Too few scenes (4))
MiniMax-M320.0%22.1%28.1%29.6%27.5%25.2%17.0%
Inkling (Xhigh)10.4%10.4%16.6%11.1%9.2%17.1%10.4% (Sample too small, Too few scenes (4))
Nemotron 3 Ultra 550B A55B (Reasoning)10.7%6.5% (Sample too small, Too few scenes (2))11.1%10.4%11.1%8.6% (Sample too small, Too few scenes (3))18.8% (Sample too small, Too few scenes (4))

Agent behavior

How each model builds its scenes: turns taken, tools called, first look at the render and use of the asset library. Averages and shares computed by Artificial Analysis from its runs.

Turns per scene
The line spans the 5th to 95th percentile, the bar the 25th to 75th, and the mark is the median. A run stops at 250 turns.
Tool calls per scene
Browser testsCodeAssetsImage viewsOther
  1. Claude Sonnet 5.5 (Max, Default Fallback)

    Turns per scene

    160

    Tool calls per scene

    185

    Render seen by turn 20: 94.3%Asset library: 83.0%

  2. Claude Opus 5.5 (Max, Default Fallback)

    Turns per scene

    137

    Tool calls per scene

    151

    Render seen by turn 20: 92.7%Asset library: 92.7%

  3. GPT-6 Astra (Max)

    Turns per scene

    60

    Tool calls per scene

    85

    Render seen by turn 20: 96.4%Asset library: 94.5%

  4. Claude Fable 5.1 (Max, Default Fallback)

    Turns per scene

    138.5

    Tool calls per scene

    139

    Render seen by turn 20: 88.9%Asset library: 88.9%

  5. GPT-6 Sol (Max)

    Turns per scene

    73

    Tool calls per scene

    103

    Render seen by turn 20: 89.1%Asset library: 81.8%

  6. GPT-6.1 Sol (Max)

    Turns per scene

    56

    Tool calls per scene

    86

    Render seen by turn 20: 94.5%Asset library: 96.4%

  7. Muse Spark 1.3 (Max)

    Turns per scene

    82

    Tool calls per scene

    86

    Render seen by turn 20: 60.0%Asset library: 20.0%

  8. GPT-6 Luna (Max)

    Turns per scene

    67

    Tool calls per scene

    77

    Render seen by turn 20: 85.5%Asset library: 32.7%

  9. Claude Haiku 5.5 (Max)

    Turns per scene

    57

    Tool calls per scene

    61

    Render seen by turn 20: 72.2%Asset library: 88.9%

  10. Grok 4.7 (Xhigh)

    Turns per scene

    81

    Tool calls per scene

    129

    Render seen by turn 20: 60.0%Asset library: 76.4%

  11. Kimi K3 (Max)

    Turns per scene

    137

    Tool calls per scene

    166

    Render seen by turn 20: 50.9%Asset library: 85.5%

  12. MiMo-V2.6-Pro

    Turns per scene

    117

    Tool calls per scene

    132

    Render seen by turn 20: 76.4%Asset library: 90.6%

  13. Gemini 3.8 Flash (High)

    Turns per scene

    109

    Tool calls per scene

    109

    Render seen by turn 20: 9.3%Asset library: 79.6%

  14. Qwen3.8 Max (0902)

    Turns per scene

    116.5

    Tool calls per scene

    129

    Render seen by turn 20: 40.0%Asset library: 83.3%

  15. GLM-5.3-Flash

    Turns per scene

    182

    Tool calls per scene

    195

    Render seen by turn 20: 47.2%Asset library: 76.9%

  16. Mistral Large 4 Preview

    Turns per scene

    122

    Tool calls per scene

    158

    Render seen by turn 20: 23.6%Asset library: 81.1%

  17. Step 5 Preview

    Turns per scene

    169

    Tool calls per scene

    180

    Render seen by turn 20: 58.2%Asset library: 85.2%

  18. GLM-5.3 (Max)

    Turns per scene

    250

    Tool calls per scene

    239

    Render seen by turn 20: —Asset library: 74.5%

  19. MiniMax-M3

    Turns per scene

    228

    Tool calls per scene

    209

    Render seen by turn 20: 7.3%Asset library: 58.2%

  20. Inkling (Xhigh)

    Turns per scene

    65

    Tool calls per scene

    70

    Render seen by turn 20: 65.5%Asset library: 27.3%

  21. Nemotron 3 Ultra 550B A55B (Reasoning)

    Turns per scene

    79.5

    Tool calls per scene

    112

    Render seen by turn 20: —Asset library: 57.7%

Tasks and scenes

The briefs Artificial Analysis published, with the scene each model built. Scenes open on Artificial Analysis's CDN in a new tab.

54 of 54 tasks

  • Eiffel TowerLandmarks and real places · 20 of 21 models have a scene

    Create an interactive Eiffel Tower in its Paris setting, recognizable both as a landmark and as an open iron structure.

  • Sagrada FamíliaLandmarks and real places · 20 of 21 models have a scene

    Create an interactive reconstruction of the Sagrada Familia in Barcelona, showing its planned complete tower composition as well as its interior architectural character. Make the reconstruction state clear.

  • Sydney Opera HouseLandmarks and real places · 21 of 21 models have a scene

    Create an interactive Sydney Opera House that reveals the relationship between its architecture, structural geometry and harbor setting.

  • Machu PicchuLandmarks and real places · 18 of 21 models have a scene

    Create an interactive Machu Picchu in its mountain landscape. Make the relationship between ridge, terraces, settlement and river gorge central to the experience.

  • Taj MahalLandmarks and real places · 19 of 21 models have a scene

    Create an interactive Taj Mahal and its garden approach, emphasizing the architectural composition, material detail and changing light.

  • Statue of LibertyLandmarks and real places · 21 of 21 models have a scene

    Create an interactive Statue of Liberty that shows both its exterior and the structure supporting its thin copper skin.

  • Sydney Airport, liveLandmarks and real places · 18 of 21 models have a scene

    Create an interactive simulation of Sydney Kingsford Smith Airport as a busy operating airport.

  • House from survey drawingsLandmarks and real places · 17 of 21 models have a scene

    Reconstruct the T. S. Eastabrook House at 200 North Scoville Avenue, Oak Park, Illinois, from the supplied survey drawings and photographs, including an explorable interior and exterior.

  • Golden Gate, 24 hoursLandmarks and real places · 15 of 21 models have a scene

    Create an accurate Golden Gate Bridge and its surroundings through a complete 24-hour day, with traffic, changing light and fog.

  • Mechanical dragonflyMachines · 17 of 21 models have a scene

    A mechanical emperor dragonfly as an exhibition piece: named selectable parts on real Odonata anatomy, four independently driven wings with phase control, a steering controller, exploded, section and schematic views, four themes and an editorial UI.

  • Rube Goldberg machineMachines · 17 of 21 models have a scene

    Create an elaborate Rube Goldberg machine that begins with a ball and ends by striking a bell. Use at least eight distinct stages drawn from mechanisms such as ramps, levers, pulleys, dominoes, a pendulum, a tipping bucket, wheels or a catapult.

  • Crash testMachines · 17 of 21 models have a scene

    Create an interactive vehicle crash test with a deforming car, deploying airbags and a restrained crash-test dummy.

  • Exploded F1 carMachines · 18 of 21 models have a scene

    Create an interactive technical demonstration of a current Formula One car, with an exploded view of its parts and a working power unit.

  • Vintage tube radioMachines · 17 of 21 models have a scene

    Create an interactive vintage vacuum-tube radio receiver with labeled internal parts, an exploded view and a working explanation of how it receives a station.

  • Watch complicationsMachines · 16 of 21 models have a scene

    Create an elegant, technically accurate mechanical watch with an exposed, labeled movement, an exploded view and a set of complications the user can select and operate.

  • Glass elevatorMachines · 18 of 21 models have a scene

    Create an interactive glass elevator in a multi-story apartment building, with a technically accurate, exposed elevator system and floors the user can choose.

  • Counterweight trebuchetMachines · 20 of 21 models have a scene

    Create an interactive counterweight trebuchet with labeled parts, in a siege where the user loads it, adjusts it and fires at a castle.

  • Galleon at seaMachines · 20 of 21 models have a scene

    Create a detailed, explorable sixteenth-century Spanish galleon sailing in changing sea conditions.

  • Dye wind tunnelNatural phenomena · 16 of 21 models have a scene

    Build an interactive wind-tunnel experiment with a two-dimensional flow field presented in a three-dimensional setting. Use at least three distinguishable dye streams to make the flow around an obstacle visible.

  • Hurricane to landfallNatural phenomena · 17 of 21 models have a scene

    Create an interactive hurricane simulation on a section of a rotating Earth, including ocean, coast, low-lying land and a populated area.

  • Carving the Grand CanyonNatural phenomena · 18 of 21 models have a scene

    Create an interactive geological time-lapse of a river carving a canyon through layered rock over millions of years.

  • Earthquake city blockNatural phenomena · 16 of 21 models have a scene

    Create an interactive earthquake experiment with at least six buildings, ranging from roughly 3 to 60 stories, responding to the same ground motion.

  • Sonic boomNatural phenomena · 19 of 21 models have a scene

    Create an interactive three-dimensional demonstration of a jet accelerating from subsonic to supersonic flight over terrain, with a ground observer.

  • Snowflake growthNatural phenomena · 20 of 21 models have a scene

    Create an interactive view of a snow crystal growing from a seed, with morphology that responds to temperature and humidity.

  • Pouring a pintNatural phenomena · 21 of 21 models have a scene

    Create an interactive beer pour with a tap, a pint glass, translucent amber liquid, rising bubbles and a developing foam head.

  • Ames room illusionNatural phenomena · 20 of 21 models have a scene

    Create an interactive Ames-room illusion: a physically distorted room that appears ordinary and rectangular from a particular viewing position.

  • Fabric on a mannequinNatural phenomena · 19 of 21 models have a scene

    Create an interactive fabric simulation in which the user dresses a mannequin in a wide range of garment fabrics, each physically simulated.

  • Inside the mitochondrionLiving bodies · 20 of 21 models have a scene

    Create an explorable journey from outside a cell to the molecular machinery inside a mitochondrion. Establish the spatial relationships clearly as the camera crosses the cell membrane, travels through cytoplasm past a microtubule, and enters the mitochondrial outer and inner membranes to reach the cristae. Stylized changes of scale are welcome if the transitions remain understandable.

  • Beating heartLiving bodies · 20 of 21 models have a scene

    Create an interactive educational model of a beating four-chamber heart, connecting electrical activation, mechanical contraction and blood flow.

  • Cat clearing a tableLiving bodies · 20 of 21 models have a scene

    Create an animated cat deliberately pushing objects off a table one at a time. Give the cat recognizable anatomy, a moving tail and a walk with planted paws that do not slide across the tabletop.

  • T. rex skeleton walkLiving bodies · 19 of 21 models have a scene

    Create an anatomically accurate, fully articulated T. rex skeleton walking through a three-dimensional scene.

  • Facial expressionsLiving bodies · 19 of 21 models have a scene

    Create an interactive human face that performs a broad range of emotions and expressive gestures, from theatrical extremes to subtle everyday expressions.

  • Weightlifter anatomyLiving bodies · 21 of 21 models have a scene

    Create an interactive adult man performing weightlifting exercises, with views that expose his muscles, skeleton and organs in anatomically accurate form.

  • Zoo habitatsLiving bodies · 18 of 21 models have a scene

    Create an explorable zoo with a wide range of animals, each in an enclosure that resembles its natural habitat, and visitors moving between them.

  • Coral reef diveLiving bodies · 20 of 21 models have a scene

    Create an interactive scuba dive over a lush coral reef, with realistic underwater light and water physics and sea animals around the diver.

  • Playable pinballGames, sports and piloting · 19 of 21 models have a scene

    Create a three-dimensional pinball table that is enjoyable to play, with a clear playfield, lighting and responsive controls.

  • Playable solitaireGames, sports and piloting · 20 of 21 models have a scene

    Create a polished, playable game of Klondike solitaire as three-dimensional cards on a table.

  • Mountain-bike trailGames, sports and piloting · 19 of 21 models have a scene

    Create an interactive mountain-bike descent through a forest trail containing berms, roots, a rock garden, a drop and a jump.

  • Pole vaultGames, sports and piloting · 18 of 21 models have a scene

    Create a realistic pole-vault simulation with an athlete whose approach speed and bar height the user controls.

  • Dodgeball matchGames, sports and piloting · 19 of 21 models have a scene

    Create an interactive dodgeball match between two teams, playable from a first-person view or watchable as a spectator.

  • Starry Night jigsawGames, sports and piloting · 19 of 21 models have a scene

    Create an interactive 500-piece jigsaw puzzle of Van Gogh's The Starry Night that can assemble itself from any state and that users can solve themselves.

  • Glider soaringGames, sports and piloting · 15 of 21 models have a scene

    Create a flyable sailplane simulation in which wind, thermals and weather shape the flight.

  • AI leaderboard steeplechaseGames, sports and piloting · 21 of 21 models have a scene

    Create an animated steeplechase race in which AI language models are the runners, with their positions taken month by month from the Artificial Analysis Intelligence Index data supplied.

  • Antennae galaxies collisionSpace · 20 of 21 models have a scene

    Create an interactive simulation of two disk galaxies colliding and merging, inspired by the Antennae galaxies, NGC 4038 and NGC 4039.

  • Kerr black holeSpace · 20 of 21 models have a scene

    Create an interactive view of a spinning black hole and its thin accretion disk, with the cinematic presence of Gargantua in Interstellar and a physically informed treatment of the light.

  • Apollo 11 missionSpace · 14 of 21 models have a scene

    Create an interactive reconstruction of Apollo 11 from launch to splashdown, combining orbital motion with close views of the spacecraft and key mission events.

  • Office on the MoonSpace · 19 of 21 models have a scene

    Create an explorable Artificial Analysis office on the Moon, with 50 people working on AI evaluation. Artificial Analysis is the company at artificialanalysis.ai that benchmarks AI models and publishes the results.

  • Starship launch and catchSpace · 19 of 21 models have a scene

    Create an animated Starship mission from launch to orbit, including the booster's return and tower catch and the ship's later landing.

  • JWST deploymentSpace · 17 of 21 models have a scene

    Create an interactive James Webb Space Telescope mission from its release by the launch vehicle through the complete unfolding of the observatory.

  • Pelican riding a bicycleStories and fantasy · 19 of 21 models have a scene

    Create an interactive scene of a recognizable pelican riding a bicycle along a course with a curve and a slope. Give the bird a long beak and throat pouch, a substantial body, webbed feet and expressive wings. It should sit on the saddle and hold the handlebars in a convincing stylized pose.

  • Octopus drum kitStories and fantasy · 20 of 21 models have a scene

    Create an octopus playing a complete drum kit. Give it a bulbous mantle, eyes and eight smoothly bending, tapering arms. Include a kick drum, snare, hi-hat, at least two toms and two cymbals arranged for an actual performance.

  • Detective murder roomStories and fantasy · 21 of 21 models have a scene

    Create an interactive crime scene from a classic detective novel: a room where a murder has occurred, with suspects present and clues hidden among the objects.

  • Dragon over a medieval townStories and fantasy · 19 of 21 models have a scene

    Create an interactive dragon attacking a medieval town with a palace and a fort, which the user can control.

  • Siege of TroyStories and fantasy · 21 of 21 models have a scene

    Create an interactive siege of Troy with both armies, accurate Late Bronze Age terrain, architecture and troops, and the Trojan Horse.

Method and versions

AA-3D v1.0 measures how well models build interactive 3D experiences. Each model works as an agent in an offline sandbox and delivers a three.js web page from a written brief; a recruited panel then compares two scenes built from the same brief without knowing which model built which.

  • AA-3D v1.0 has 55 tasks across seven domains (real places, machines, natural phenomena, living bodies, games, space, fiction); each is the brief for one interactive scene.
  • Each model runs in Stirrup, Artificial Analysis' open-source harness, in a sandbox without network access that has Blender, Rapier physics, Python, a headless browser and a CC0 asset library; a run stops after three hours or 250 turns.
  • A recruited panel votes blind between two scenes built from the same brief.
  • Elo and its 95% confidence interval come from those votes; GPT-6 Luna (Max) is fixed at 1,000.
  • Cost and tokens per scene use Artificial Analysis' data and rules over the 55 tasks: cache reads at the cache price, other input at the cache-write price, output at the output price. A model without a cache rate or a price has no cost.

Data via Artificial Analysis, AA-3D v1.0.

Preferences and analytics

Rybbit: visits, performance, clicks, copies, forms and session replay (masked inputs), with your permission. Refusing does not limit the site.

Privacy

Allow navigation analytics?

With your permission, Rybbit measures visits, performance and interactions (clicks, copies, forms), including session replay, to improve BenchSift. Input values are masked in replays. Refusing does not limit the site.

Your choice is stored for 180 days. Change it at any time using “Privacy preferences” in the footer. Withdrawal reloads the page.

No choice stored: Rybbit is disabled.

Privacy