An open model leaderboard

Every AI model. Every task.
Ranked.

Name: GoldieBench — one-shot AI model build benchmark
Creator: Julian Goldie
License: https://creativecommons.org/licenses/by/4.0/

The Goldie AI Model Leaderboard — every model, every task, every demo, ranked. Click into any cell for the live, playable demo — every model gets the same prompt, every result is on the same page.

1310Live demos

50Tasks

33Models

1041Curated verdicts

The leaderboard

Ranking = Julian's actual 0–10 scores averaged across all tasks both ran. Medals = per-task rank (highest = 🥇 gold, second = 🥈 silver, third = 🥉 bronze). Models without scored verdicts are unranked — their demos are still on the bench, but Julian hasn't put a number on them yet.

Scores extracted from Julian's GLM-5.2 vs Kimi K2.7 vs Opus 4.8 and GLM-5.2 vs Qwen 3.7 vs Opus 4.8 guides. See methodology for data provenance.

Ranked by overall average across all tasks. Tap a category to see who wins at it.

Click any column header to sort ↕

Rank	Model	Best at	Tasks	Medals	Avg / 10	Reference benchmark
#1	Fusion OpenRouter	★ best at Games	47 tasks47/47 scored	🥇 21🥈 3🥉 1	8.59	69.0% DRACO deep-research (premium panel)
#2	Claude Opus 5 Anthropic	★ best at Sims	50 tasks50/50 scored	🥇 13🥈 7🥉 5	8.27	—
#3	Hermes MoA Hermes · Mixture of Agents	★ best at Visuals	47 tasks47/47 scored	🥇 3🥈 10🥉 2	8.17	—
#4	GPT-5.6 Sol OpenAI	★ best at Pages	50 tasks50/50 scored	🥇 2🥈 9🥉 10	8.16	60.5 (+8.7 vs 5.5) HealthBench Professional (Sol)
#5	Claude Fable 5 Anthropic	★ best at Games	47 tasks47/47 scored	🥇 4🥈 2🥉 1	8.10	7.72 avg One-shot pass, before agentic repair of 8 crashed/weak tasks
#6	Qwen 3.8 Alibaba	★ best at Sims	45 tasks45/45 scored	🥇 8🥈 6🥉 7	8.10	—
#7	Grok xAI	★ best at Games	47 tasks43/47 scored	🥇 5🥈 1🥉 1	8.09	—
#8	MiniMax M3 MiniMax	★ best at Games	47 tasks47/47 scored	🥇 2🥈 1🥉 3	7.97	1,048,576 tokens Context window
#9	Fugu Ultra Sakana AI	★ best at Visuals	42 tasks42/42 scored	🥇 5🥈 2🥉 2	7.94	73.7 SWE Bench Pro
#10	Kimi K3 Moonshot AI	★ best at Games	50 tasks50/50 scored	🥇 8🥈 3🥉 7	7.89	~Fable/Sol tier; Terminal Bench the standout Launch-day community benchmarks
#11	GLM-5.2 Zhipu / Z.ai	★ best at Pages	47 tasks47/47 scored	🥇 5	7.77	62.1% SWE-bench Pro
#12	Fugu Mini Sakana AI	★ best at Visuals	37 tasks36/37 scored	🥇 2	7.75	—
#13	Claude Opus 5.5 Anthropic	★ best at Sims	50 tasks50/50 scored	🥉 1	7.57	—
#14	Muse Spark 1.2 Meta	★ best at Sims	50 tasks50/50 scored	🥈 3🥉 4	7.55	—
#15	Opus 4.8 Anthropic	★ best at Sims	47 tasks47/47 scored	🥇 3🥈 1	7.51	88.6% SWE-bench Verified
#16	Kimi K2.7 Moonshot AI	★ best at Games	47 tasks25/47 scored	🥇 1🥈 2	7.46	80.2% SWE-bench Verified (K2.6 — K2.7 internal)
#17	Grok 4.7 xAI	★ best at Games	20 tasks20/20 scored	🥈 2🥉 3	7.15	46 Artificial Analysis Intelligence Index v4.3.2
#18	Gemini 3.6 Flash Google	★ best at Pages	50 tasks50/50 scored	🥇 1🥈 2🥉 1	7.08	—
#19	Claude Sonnet 5 Anthropic	★ best at Pages	47 tasks47/47 scored	🥉 4	7.01	82.1% SWE-bench Verified
#20	Qwen 3.7 Alibaba	★ best at Games	47 tasks47/47 scored	—	7.00	—
#21	Fugu Ultra 1.1 Sakana AI	★ best at Games	24 tasks23/24 scored	🥈 1🥉 1	6.94	73.7 SWE-Bench Pro (reported)
#22	MiMo-V2.6 Pro Xiaomi	★ best at Pages	50 tasks50/50 scored	🥇 5🥈 8🥉 4	6.35	46.32 Artificial Analysis Intelligence Index v4.3
#23	Inkling Thinking Machines	★ best at Pages	50 tasks50/50 scored	—	6.07	77.6% — ahead of Nemotron 3 Ultra (70.7) SWE-bench Verified (agentic coding)
#24	Grok 4.6 xAI	★ best at Games	20 tasks20/20 scored	🥇 1	5.95	61 Artificial Analysis Intelligence Index
prov	LongCat-2.0 Meituan	★ best at Games	4 tasks4/4 scored · provisional	—	8.12	70.8 Terminal-Bench 2.1
prov	Hy3 Tencent Hunyuan	★ best at Games	7 tasks7/7 scored · provisional	—	6.76	2.67/4 Expert blind eval
—	DeepSeek V4 Flash DeepSeek	—	50 tasks0/50 scored	—	unranked	—
—	DeepSeek V4 Pro DeepSeek	—	50 tasks0/50 scored	—	unranked	—
—	Kimi K2.7 · Fast Moonshot AI	—	47 tasks0/47 scored	—	unranked	65s Build time · solar system (one shot)
—	Kimi K2.7 · No-Think Moonshot AI	—	47 tasks0/47 scored	—	unranked	47s Build time · solar system (one shot)
—	Kimi K2.7 · Quality Moonshot AI	—	47 tasks0/47 scored	—	unranked	71s Build time · solar system (one shot)
—	Claude Mythos 5 Anthropic	—	0 tasks	—	unranked	Vetted partners only Restricted access
—	Kilo Code Kilo	—	0 tasks	—	unranked	8.3/10 Cheaper-plan quality (GPT-5.5 cohort)

The matrix — tasks × models

Each cell is a live, playable demo. Click any thumbnail to run that model's attempt at that task. Empty cells mean we haven't tested that combo yet.

Task ↓ · Model →

Game

Crypt

🥇

🥈

🥉

—

—

Game

Dragonrealm

🥇

🥉

🥇

🥉

🥉

🥉

Game

Doom

🥈

🥇

▶ PLAY

🥈

🥈

—

Game

Racing

🥇

🥉

🥇

🥉

—

Game

Skyrim

🥇

🥇

🥉

—

Game

Twilightvale

🥇

🥉

🥇

🥉

🥉

—

—

Game

Voxelcraft

🥇

🥇

—

🥉

—

—

—

Game

Gtafoot

🥇

🥈

—

🥉

—

—

Game

Gtadrive

🥇

🥉

🥉

🥈

—

🥉

—

—

Page

Aipbpromo

🥈

🥈

🥇

—

—

—

—

Game

Parachute

🥇

🥈

—

—

—

🥈

—

Game

Flightsim

🥇

🥈

—

—

🥉

—

Game

Arcade

🥈

🥈

🥈

🥈

🥇

—

Game

Dogfight

🥇

🥈

🥈

🥈

—

Game

Neonracer

🥇

🥈

🥈

—

Game

Nordiccrypt

🥇

🥈

🥈

—

Game

Outrun

🥇

🥇

🥇

🥇

—

Game

Rpg

🥈

🥉

🥇

—

Game

Neonblaster

🥇

🥈

—

🥈

—

Game

Neoncity

—

🥈

🥇

🥈

🥈

—

Game

Raycaster

🥈

—

🥈

🥇

🥈

—

Visual

Aurora

🥈

🥈

🥈

🥈

🥈

🥇

🥈

—

—

Game

Dragonflight

🥇

🥈

🥉

🥉

🥉

—

—

Game

🥇

🥇

🥇

🥇

—

—

Sim

Blackhole

🥇

🥇

🥇

—

—

🥇

—

Sim

Boids

🥇

🥇

🥇

🥇

—

—

🥇

—

Sim

Cloth

🥉

🥈

🥇

—

—

—

Visual

Fireworks

🥇

🥇

🥇

—

—

—

Sim

Fluid

🥉

🥇

🥇

—

🥉

—

—

Sim

Fractal

🥈

🥈

🥈

🥇

—

—

—

Sim

Galaxy

🥇

🥉

🥈

🥉

—

▶ PLAY

🥉

—

🥉

—

Page

Landing

🥇

🥇

🥇

🥇

🥇

—

—

—

Visual

Lavalamp

🥇

🥇

—

—

🥇

—

Visual

Matrix

🥈

🥈

🥈

🥈

—

🥇

—

🥈

—

Sim

Orbit

🥇

🥉

🥉

🥉

🥇

—

—

🥉

—

Sim

Particleforge

🥈

🥇

🥉

🥉

🥉

—

—

—

Visual

Plasma

🥇

🥈

🥈

—

🥈

—

—

Sim

Reactiondiff

🥈

🥇

—

—

🥈

—

Sim

Solar

🥇

🥉

🥉

🥉

🥉

🥇

—

—

—

Visual

Synthwave

🥉

🥉

🥇

🥇

—

—

—

Visual

Terrain

🥈

🥇

—

🥈

—

🥈

—

Visual

Voxel

🥇

🥇

🥇

—

—

—

Visual

Waves

🥇

🥇

—

—

🥇

—

Page

Webos

🥇

🥉

🥇

🥉

—

—

🥉

—

Sim

Wormhole

🥇

🥈

🥉

🥉

🥉

—

—

🥉

—

Sim

Pathtracer

🥇

🥇

🥉

🥉

—

🥉

—

—

Game

Pool

🥇

🥈

—

—

🥉

—

🥉

—

Other

Matrixrain

—

🥇

—

🥉

—

🥉

—

—

—

🥇

—

—

Other

Mlx Speedtest

—

🥈

—

🥇

—

—

—

—

🥈

—

—

Other

Neonsnake

—

🥈

—

🥉

—

—

🥇

—

—

—

—

—

Showing the 50 most-tested tasks. See all 50 tasks → · Browse all 66 head-to-head comparisons →

The models on the bench

33 AI models tested so far — every frontier release I've put through the same fixed prompt set inside the Agent Operating System. Click any card for the full model review, every demo, and direct head-to-head comparisons. Models without published Julian-graded per-task scores (Fable 5, Mythos 5, Kilo) are listed as unranked until I publish their scored head-to-head guides. Fusion is currently graded by Claude's AI judge (read each demo's HTML, compared to existing scored attempts on the same prompt) — its scores carry a "claude-bench-judge" source label and will be replaced with Julian's own scores when his Fusion head-to-head ships.

Fusion OpenRouter

Multi-model panel — Fable 5 + GPT-5.5, ensembled. Beats Fable 5 at half the price.

8.59avg

47tasks

21🥇

3🥈

Claude Opus 5 Anthropic

The new Anthropic flagship — benched on all 45 one-shot builds the day it landed.

8.27avg

50tasks

13🥇

7🥈

Hermes MoA Hermes · Mixture of Agents

A panel of frontier models, merged by a chair. The model doesn't matter — the system does.

8.17avg

OpenAI's flagship — the Sun of the 5.6 lineup.

8.16avg

50tasks

2🥇

9🥈

Claude Fable 5 Anthropic

The newest Anthropic model — first Mythos-class made generally available.

8.10avg

Alibaba's 2.4T flagship — benched through Qoder.

8.10avg

Snappy + real-time — the X-native model.

8.09avg

1M-context frontier model at $0.30/M tokens — cheapest big-context model on the bench.

7.97avg

Sakana's multi-agent answer to Fusion — frontier ensemble without single-vendor risk.

7.94avg

Moonshot's 2.8T flagship — 1M context, tuned for long-horizon agent work.

7.89avg

The never-forgets agent — 1M context, open weights.

7.77avg

Fugu's fast mini variant — single model, no panel, ~3 min per build.

7.75avg

37tasks

2🥇

0🥈

Claude Opus 5.5 Anthropic

Anthropic's Opus 5.5 — benched on all 50 one-shot builds, every game playtested.

7.57avg

Meta's coding reasoning model — co-trained with its own agent, 1M-token window.

7.55avg

The reasoning king — deepest thinking, premium price.

7.51avg

47tasks

3🥇

1🥈

Kimi K2.7 Moonshot AI

The heavy lifter — frontier coder at flat-rate.

7.46avg

Grok 4.6's price, a bigger base model, and it builds games that hold together.

7.15avg

20tasks

0🥇

2🥈

Gemini 3.6 Flash Google

Google's launch-day Flash — faster, cheaper, fewer tokens.

7.08avg

50tasks

1🥇

2🥈

Claude Sonnet 5 Anthropic

The agentic SWE frontier — 82% SWE-bench Verified, Dev Team mode.

7.01avg

Multilingual open-weights — strong on Chinese reasoning.

7.00avg

47tasks

0🥇

0🥈

Fugu Ultra 1.1 Sakana AI

Sakana's multi-agent orchestrator, v1.1 — routes experts per request.

6.94avg

Open weights that score level with Opus 5 on agents, for cents.

6.35avg

50tasks

5🥇

8🥈

Inkling Thinking Machines

A 975B open-weights frontier model — yours to own and run.

6.07avg

Frontier intelligence at half the frontier price.

5.95avg

The open 1.6T MoE that builds — a frontier coder trained on non-Nvidia ASIC superpods.

8.12avg

Tencent's open-weights coder — Apache-2.0, cheap, beats GLM-5.1 on frontend in Tencent's blind eval.

6.76avg

7tasks

0🥇

0🥈

DeepSeek V4 Flash DeepSeek

DeepSeek's cheap tier, retrained for agents — same size, sharper loops.

DeepSeek V4 Pro DeepSeek

DeepSeek's flagship tier — benched head-to-head against its own cheap Flash.

Kimi K2.7 · Fast Moonshot AI

Fast mode — top speed, minimal thinking.

Kimi K2.7 · No-Think Moonshot AI

Pure execution mode — no chain of thought.

Kimi K2.7 · Quality Moonshot AI

Quality mode — deepest thinking, best output.

Claude Mythos 5 Anthropic

Restricted-access flagship — vetted partners only.

Fable 5-class intelligence at ~59% less. The split-the-cost play.

What is Goldie Bench?

Goldie Bench is a one-shot AI model leaderboard. Every model gets the same fixed prompt, single HTML file out, live on the page. No iteration. No "best of N." What came out on the first run is what's on the matrix — and I score each result 0–10 on whether it ran, how close it hit the brief, and how good it looked.

Who I am — and why I built this.

Julian Goldie — Founder, Goldie Agency + AI Profit Boardroom + Julian Goldie SEO YouTube

I'm Julian Goldie. I run Goldie Agency — a 7-figure SEO + link-building agency with 70+ employees that ships AI-driven work for clients every day. I built the Agent Operating System — a dashboard where my AI crew dispatches frontier models from one shared kanban — so I can run it all. The bench is a slice of my actual daily work, not a synthetic benchmark.

I also run the AI Profit Boardroom (3,400+ founders, 38 countries), publish daily AI + SEO tutorials on my YouTube channel (400k subscribers), post on X (163k followers), and wrote Link Building Mastery on Amazon. Every model on this leaderboard is one I dispatched personally from inside Agent OS.

▶ YouTube · 400k 𝕏 X · 163k in LinkedIn Skool · AIPB Goldie Agency Full About →

Why I built this AI model leaderboard

Standard AI benchmarks — MMLU, HumanEval, SWE-bench, LiveCodeBench — measure narrow things. Did the model output text that matches an expected answer? Did the code compile? Did the patch pass the test? These benchmarks are useful for vendors who want to make a number go up. They have almost nothing to do with whether a frontier model can ship a thing on the first prompt.

The buyer audience for these models — founders, builders, indie devs, agencies, anyone running AI agents day to day — needs a different question answered: can you ask the model for a playable game, a working simulation, a deployable page, and have it ship in one prompt? That's the test that decides whether you'd wire the model into your stack.

Goldie Bench is that test. Same fixed prompt for every model. Single HTML out. Live on the page. Real 0–10 score from me, posted publicly with the reasoning. No vendor pays me. No score gets buried.

How I test — the methodology in 60 seconds

I pick a creative coding prompt a frontier AI model should be able to ship in one shot — raycaster maze, fluid sim, neon city flythrough, top-down RPG, landing page.
I dispatch the exact same prompt to each model from the kanban inside the Agent Operating System. No system-prompt cheats, no few-shot tricks.
I save whatever .html file the model produces on the first run. No iteration. No coaching.
I score each result 0–10 on three axes: did it run, did it hit the brief, did it look good.
I publish the scores publicly in the source comparison guides on agentos.guide — and that's the data behind every cell on this page.

Full data provenance on the methodology page.

What makes Goldie Bench different from other AI model leaderboards

Live demos, not numbers in a table. Click any cell on the matrix and the model's actual one-shot HTML opens in a new tab. You don't have to trust my score — you can play the result and form your own opinion.
Real 0–10 scores from a working operator, not auto-evaluated against a reference. I run the models inside the same Agent OS my 3,400+ community uses daily, and I score on what would actually ship.
Same prompt across every model — no per-vendor handicapping. Every model sees the exact same string.
Publicly sourced data. Every score traces back to a published comparison guide on agentos.guide. Cells without a score are honestly marked as unranked.
Updated daily-ish. When a new model ships, I run the same prompts through it on camera (you can watch on YouTube, 400k+ subscribers) and the bench updates.

FAQ — about the bench

What's the best AI model right now?

Across 1041 scored cells from my source guides, Opus 4.8 tops the leaderboard at 8.46/10 average — most consistent across game-feel, raycasters, and one-shot reasoning. GLM-5.2 is the open-weights challenger at 8.23/10 with the prettiest visual builds. Pick by use case — see Opus 4.8 vs GLM-5.2 for the head-to-head.

Is Goldie Bench the same as the LMArena, Hugging Face, or Chatbot Arena leaderboard?

No. Those rank models on human-preference Elo votes or quiz-style benchmarks. Goldie Bench ranks on a single fixed creative-coding task per cell — same prompt, one shot, scored 0–10 by me. Different methodology, different answer.

How often do new models get added?

Whenever a new frontier release lands. I usually wire it into the Agent OS the same week, run the fixed prompt set through it on camera, and the bench updates. Recent additions: GLM-5.2, Kimi K2.7, Qwen 3.7.

Can I add a model or task?

Ideas land best inside the AI Profit Boardroom community. Members propose tasks, I test, scores get added to the bench.

Are these scores affiliate-driven or vendor-sponsored?

No. No vendor pays me. No score is gated. The Agent OS that runs these tests is built and maintained by me; the community that powers the workflow lives inside AIPB. That's the only commercial layer on this site.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders

258documented wins

38countries

$69/momonthly

Join AIPB · $69/mo → Read the Agent OS guides →

Every AI model. Every task.Ranked.

The leaderboard

The matrix — tasks × models

The models on the bench

What is Goldie Bench?

Who I am — and why I built this.

Why I built this AI model leaderboard

How I test — the methodology in 60 seconds

What makes Goldie Bench different from other AI model leaderboards

FAQ — about the bench

What's the best AI model right now?

Is Goldie Bench the same as the LMArena, Hugging Face, or Chatbot Arena leaderboard?

How often do new models get added?

Can I add a model or task?

Are these scores affiliate-driven or vendor-sponsored?

Run this stack yourself.

Every AI model. Every task.
Ranked.