BuilderBuilder
Reproduction library

Open the experiment. Rebuild the evidence.

Browse each benchmark once. Stages, prompt versions, and supplemental workflows stay listed under that benchmark. Released Builder downloads stay on the tests that include them.

Browse all 28 benchmarksVerify the benchmark evidence
28benchmark families in this catalog, shown once each
Benchmark families

28 benchmark families

12 families include at least one released Builder download. 16 families are public evidence. Variant runs, prompt receipts, and material gaps stay on the benchmark page.

Rubik's cube, single turn public evidence6 variants

Rubik’s Cube

  • 6 variants
  • 25 runs
  • 12 related records
  • No Builder download

Rubik’s Cube single-turn and three-turn tests. Each stage and workflow keeps its own prompt, runs, and status.

Open benchmark
Blossom arena public evidence2 variants

Blossom arena

  • 2 variants
  • 16 runs
  • 3 related records
  • No Builder download

Blossom arena tests. Each prompt hash keeps its own runs and status.

Open benchmark
Western frontier town with a player character and saddled horse in Red Sands v12 variants

Red Sands

  • 2 variants
  • 5 runs
  • 1 released material set

Red Sands tests. The v1 release and ledger version v2 keep their own prompt records, runs, and status.

Open benchmark
Robot pick-and-place public evidence3 variants

Robot pick-and-place

  • 3 variants
  • 15 runs
  • 2 related records
  • No Builder download

Robot pick-and-place tests for the apple-carry task. Unaided runs, the decision loop, and the Astra planner plus Jev workflow keep their own records. Apple stem is a separate family.

Open benchmark
Eiffel Tower drawing public evidence3 variants

Eiffel Tower drawing

  • 3 variants
  • 14 runs
  • 7 related records
  • No Builder download

Eiffel Tower drawing tests. Unaided runs, the decision loop, and the Astra planner plus Jev workflow keep their own prompt records, runs, and status.

Open benchmark
Apple stem public evidence4 variants

Apple stem

  • 4 variants
  • 15 runs
  • 7 related records
  • No Builder download

Apple stem tests. Unaided runs, the decision loop, the Astra planner plus Jev workflow, and the Astra Jev hybrid keep their own records.

Open benchmark
Two-handed puzzle public evidence2 variants

Two-handed puzzle

  • 2 variants
  • 7 runs
  • 3 related records
  • No Builder download

Two-handed puzzle tests. Unaided runs and the decision loop keep their own prompt records, runs, and status.

Open benchmark
Spacecraft approaching a ringed planet in the Space Flight v1 collection2 variants

Space Flight

  • 2 variants
  • 35 runs
  • 11 related records
  • 1 released material set

Space Flight tests. Stage 1 and Stage 2 — Landfall keep their own prompts, runs, and status.

Open benchmark
Infinite cathedral shader output in the Shader Frontier v1 collection1 variant

Infinite Cathedral

  • 1 variant
  • 21 runs
  • 10 related records
  • 1 released material set

Infinite Cathedral. The published test keeps its own prompt, runs, and status.

Open benchmark
Neo-Gothic Storm City benchmark output1 variant

Neo-Gothic Storm City

  • 1 variant
  • 20 runs
  • 12 related records
  • 1 released material set

Neo-Gothic Storm City. The published test keeps its own prompt, runs, and status.

Open benchmark
Campfire benchmark output in the Campfire Interactive Scene v1 collection1 variant

Campfire Under a Starry Night

  • 1 variant
  • 18 runs
  • 7 related records
  • 1 released material set

Campfire Under a Starry Night. The published test keeps its own prompt, runs, and status.

Open benchmark
Voxel jungle-temple diorama in the Blender Voxel Worlds v1 collection1 variant

Jungle Temple

  • 1 variant
  • 15 runs
  • 7 related records
  • 1 released material set

Jungle Temple. The published test keeps its own prompt, runs, and status.

Open benchmark
Shrine Village benchmark output1 variant

Shrine Village

  • 1 variant
  • 9 runs
  • 5 related records
  • 1 released material set

Shrine Village. The published test keeps its own prompt, runs, and status.

Open benchmark
Oasis Outpost benchmark output1 variant

Oasis Outpost

  • 1 variant
  • 16 runs
  • 8 related records
  • 1 released material set

Oasis Outpost. The published test keeps its own prompt, runs, and status.

Open benchmark
Silver laptop illuminated in the MacBook Cinematic v1 collection1 variant

MacBook Cinematic

  • 1 variant
  • 9 runs
  • 2 related records
  • 1 released material set

MacBook Cinematic. The published test keeps its own prompt, runs, and status.

Open benchmark
Knight facing an armored boss in the JRPG Boss Battle v1 collection1 variant

JRPG Boss Battle

  • 1 variant
  • 22 runs
  • 3 related records
  • 1 released material set

JRPG Boss Battle. The published test keeps its own prompt, runs, and status.

Open benchmark
Arena-only preview from the clean Fighter Game v1 result1 variant

Fighter Game

  • 1 variant
  • 7 runs
  • 1 released material set

Fighter Game. The published test keeps its own prompt, runs, and status.

Open benchmark
Elevated night view over the Shrine Exploration courtyard with a samurai1 variant

Shrine Exploration

  • 1 variant
  • 1 run
  • 1 released material set

Shrine Exploration. The published test keeps its own prompt, runs, and status.

Open benchmark
ATS depot driving public evidence1 variant

ATS depot driving

  • 1 variant
  • 3 runs
  • No Builder download

ATS depot driving. The published test keeps its own prompt, runs, and status.

Open benchmark
Bluetooth speaker PCB public evidence1 variant

Bluetooth speaker PCB

  • 1 variant
  • 3 runs
  • 3 related records
  • No Builder download

Bluetooth speaker PCB. The published test keeps its own prompt, runs, and status.

Open benchmark
CAD-002 robot arm public evidence1 variant

CAD-002 robot arm

  • 1 variant
  • 3 runs
  • No Builder download

CAD-002 robot arm. The published test keeps its own prompt, runs, and status.

Open benchmark
Doom with Jev public evidence1 variant

Doom with Jev

  • 1 variant
  • 2 runs
  • No Builder download

Doom with Jev. The published test keeps its own prompt, runs, and status.

Open benchmark
Interactive voxel viewer public evidence1 variant

Interactive voxel viewer

  • 1 variant
  • 13 runs
  • 5 related records
  • No Builder download

Interactive voxel viewer. The published test keeps its own prompt, runs, and status.

Open benchmark
Neon Gunner public evidence1 variant

Neon Gunner

  • 1 variant
  • 15 runs
  • 3 related records
  • No Builder download

Neon Gunner. The published test keeps its own prompt, runs, and status.

Open benchmark
Off-road Mud Game public capture1 variant

Off-road Mud Game

  • 1 variant
  • 11 runs
  • 1 related record
  • No Builder download

Off-road Mud Game. The published test keeps its own prompt, runs, and status.

Open benchmark
Sekiro main menu public evidence1 variant

Sekiro main menu

  • 1 variant
  • 16 runs
  • 3 related records
  • No Builder download

Sekiro main menu. The published test keeps its own prompt, runs, and status.

Open benchmark
Snowboarding public evidence1 variant

Snowboarding

  • 1 variant
  • 8 runs
  • 3 related records
  • No Builder download

Snowboarding. The published test keeps its own prompt, runs, and status.

Open benchmark
Turbofan CAD public evidence1 variant

Turbofan CAD

  • 1 variant
  • 8 runs
  • No Builder download

Turbofan CAD. The published test keeps its own prompt, runs, and status.

Open benchmark
Public benchmark evidenceVerify what happened.

Results, costs, tokens, workflow time, run ledgers, methodology, evidence, and known gaps remain free to inspect.

Free public comparisons

Latest published comparisons

Inspect the published outcomes first. Evidence details and known gaps stay accessible below each comparison.

I Tested Opus 5.5 vs GPT-6 Astra, Sonnet 5.5, GPT-6.1 SolLaunch 006 · public comparison
86 recorded runsPublished 2026-10-08

Opus 5.5 vs GPT-6 Astra vs Sonnet 5.5 vs GPT-6.1 Sol

Eiffel Tower drawing · Flooded neo-gothic city · Infinite cathedral · Turbofan CAD · Bluetooth speaker PCB · Blossom arena · Space flight · Snowboarding · Doom with Jev · Oasis outpost · Jungle temple · Robot pick-and-place · Apple stem · Two-handed puzzle · Rubik's cube, single turn · Rubik's cube, three turns · JRPG boss battle · Neon Gunner · Sekiro main menu · ATS depot driving · Supplemental · CAD-002 robot arm · Supplemental · Eiffel drawing, Astra planner plus Jev · Supplemental · Apple carry, Astra planner plus Jev

Evidence available & known gaps

Evidence available

  • Published episode comparison and official video
  • Ledger status kept separate from episode narration
  • Hash-bound public footage only where the clip key matches the attempt

Still missing

  • No aggregate score or overall winner
  • No new blind-vote cohort
  • Astra planner + Jev clips are supplemental and are not unaided episode attempts
  • Repeat-run reliability evidence
I Tested Qwen 3.8 Flash vs Deepseek V4.1 Flash, GLM 5.3 Flash, Gemini 3.8 FlashLaunch 005 · public comparison
54 recorded runsPublished 2026-09-18

Qwen 3.8 Flash vs DeepSeek V4.1 Flash vs GLM 5.3 Flash vs Gemini 3.8 Flash

Rubik’s Cube single turn · Rubik’s Cube three-move extension · Campfire · Courtyard Blossom Stage 2 / Blender arena · Interactive voxel diorama · JRPG boss battle · Neon Gunner · Sekiro main menu · Space Flight Stage 1 · Turbofan CAD · Neo-Gothic Storm City · Infinite Cathedral · Eiffel Tower drawing · Apple-stem manipulation

Evidence available & known gaps

Evidence available

  • Fifty-four exact direct clip-to-result mappings across fourteen fixtures
  • Thirty-five reused public records plus nineteen explicit campaign-evidence records
  • Committed CAD and robotics completion boundaries
  • Published editorial assessment and full episode link
  • Model, route, reasoning, validation, time, cost, and caveat disclosures where recorded

Still missing

  • Repeat-run reliability evidence
  • Any justified aggregate score or universal winner
Samurai looking up through the illuminated torii in Shrine ExplorationLaunch 004 · public comparison
32 recorded runsPublished 2026-07-31

Claude Opus 5 across nine game-building tests

Off-road Mud Game · Space Flight Stage 2 — Landfall · Neo-Gothic Flooded City · Infinite Cathedral · MacBook Pro · JRPG Boss Battle · Jungle Temple · Oasis Outpost · Shrine Exploration

Evidence available & known gaps

Evidence available

  • Cost, workflow time, token, stack, artifact, validation, and caveat ledgers
  • Nine frozen tests with 32 exact featured records
  • Fourteen hash-bound Launch 004 presentation recordings
  • Public Shrine receipts, disclosed prompt privacy findings, platform limitations, and exclusion proof
  • Published owner benchmark assessment

Still missing

  • Companion video episode metadata (deferred after benchmark publication)
  • Repeat-run reliability evidence
  • Complete Shrine subagent token mix and exact total cost
  • WebGL gameplay, physical-input, and audible-output verification
  • Windows runtime, gameplay, physical-input, and audio verification
Silver laptop illuminated against a dark studio backgroundLaunch 003 · public comparison
42 recorded runsPublished 2026-07-23

I Gave Qwen3.8 Max, Kimi K3, GPT-5.6 and Fable 5 the Same 9 Game Tests

Neo-Gothic Storm City · MacBook-Class Cinematic · JRPG Boss Battle · Campfire Under a Starry Night · Space Flight Game · Shrine Village · Infinite Cathedral · Oasis Outpost · Jungle Temple

Evidence available & known gaps

Evidence available

  • Cost, workflow time, token, stack, artifact, and caveat ledgers
  • Nine frozen Qwen tests with a 42-record comparison pool
  • Nine public Qwen media recordings and linked artifacts
  • Published owner editorial assessment and official video

Still missing

  • Minimum community preference sample
  • Repeat-run reliability evidence
  • A completed Qwen MacBook agent run
Knight facing an armored boss in a volcanic arenaLaunch 002 · public comparison
33 recorded runsPublished 2026-07-19

Kimi K3 vs GPT-5.6 SOL vs Claude Fable 5: 9 Game Tests

Neo-Gothic Storm City · MacBook-Class Cinematic · JRPG Boss Battle · Campfire Under a Starry Night · Space Flight Game · Shrine Village · Infinite Cathedral · Oasis Outpost · Jungle Temple

Evidence available & known gaps

Evidence available

  • Cost, workflow time, token, stack, and artifact ledgers
  • Nine frozen tests with 33 featured run records
  • Seven anonymous Kimi-versus-Fable blind matchups
  • Published editorial assessment and official video

Still missing

  • Minimum community preference sample
  • Repeat-run reliability evidence
  • One reconciled Kimi MacBook API request
Spacecraft approaching a ringed planet through an asteroid fieldLaunch 001 · public comparison
28 recorded runsPublished 2026-07-15

GPT-5.6 vs Claude Fable 5: Same Prompts, 7 AI Game-Dev Tests

Infinite Cathedral Corridor · Neo-Gothic Storm City · Space Flight Game · Campfire Under a Starry Night · Jungle Temple · Shrine Village · Oasis Outpost

Evidence available & known gaps

Evidence available

  • Cost, workflow time, token, stack, and artifact ledgers
  • Seven standardized Fable-versus-Sol comparisons
  • Anonymous blind exploration; sign-in only to record a vote
  • Published editorial assessment and official video

Still missing

  • Minimum community preference sample
  • Repeat-run reliability evidence
Community access is free

Discuss the result. Build the next one.

Join after signing in. Builder adds released reproduction components and private Builder Lab; it is not required for the community Discord or public benchmark evidence.

Join the verified member Discord →