Get the Agent OS + join 4,000+ founders inside the AI Profit Boardroom → Join AIPB ($59/mo)

GoldieBench deep dive

Fusion's full benchmark breakdown.

Multi-model panel — Fable 5 + GPT-5.5, ensembled. Beats Fable 5 at half the price. Every number below is either one of my own judged one-shot builds - playable on this site - or an outside result with its source linked. Nothing display-only, nothing vibes.

Data refreshed
2026-07-17
Scored tasks
47
Reading time
10 min
External sources
2

01 · The headline numbers

GoldieBench average
8.59/10
47 scored one-shot tasks
Board rank
#1
of 17 ranked frontier models
Task medals
22🥇 4🥈 8🥉
outright wins on shared briefs
Context
Varies — depends on which panel models are dispatched
OpenRouter Fusion API pricing

02 · Every benchmark, bar by bar

All 47 scored tasks, best first. Hover for the judge's comment; click through for every model's build on that task. Or overlay other models in the interactive graphs.

03 · Where it wins - with the judge's own words

“RETRY @ 24K tokens — now complete: 29KB with rAF + 7 input handlers + closed tags. Torch-lit ancient ruin, PointerLockControls, bloom, chasing enemies, boss room. The original truncated attempt has been replaced with a working build.”

- the judge's verdict, unedited

“RETRY @ 24K tokens — now complete: 44KB three.js + WebGL using renderer.setAnimationLoop (three.js native loop), 6 input handlers, full update() per-frame: player + NPCs + enemies + weather + day/night + HUD. Densest build on the bench.”

- the judge's verdict, unedited

“Photon back-tracing through a curved-space metric (claims as much in the code) for actual gravitational lensing — disk's far side lifted over and under the shadow. Loading screen says "computing space-time metric…" Range-slider parameter panel for spin/disk tilt/exposure. Most ambitious blackhole attempt by code structure.”

- the judge's verdict, unedited

“First-person Nordic dungeon on three.js with PointerLockControls + WebGL. Torch-lit corridors, held torch, skeletons to strike, health + gold HUD. The crypt Julian wanted.”

- the judge's verdict, unedited

“RETRY @ 24K tokens — now complete: 36KB three.js + WebGL with 2 rAFs, 6 input handlers, full </html>. Fighter, enemies, missiles, guns, sky + ground, full HUD. The original truncated 32KB attempt has been replaced with a working build.”

- the judge's verdict, unedited

04 · Where it struggles - quoted, not hidden

Boards that hide the weak rows are brochures. These are Fusion's lowest scored builds, verdicts unedited:

“WebGL metaball shader with warm gradient. Smooth blob morphing, low-frequency motion. Simpler than the deep builds but on brief.”

- the judge's verdict, unedited

“Aurora ribbons over a mountain ridge silhouette, WebGL noise shader. Slow morphing curtains in cyan + magenta. Smaller scope (10KB) but the prompt is met.”

- the judge's verdict, unedited

“120KB · auto-plays clean (video ad, input n/a) · css timeline”

- the judge's verdict, unedited

05 · Category breakdown vs the whole field

Games (23 tasks)

Fusion8.8
field avg7.1

Others (0 tasks)

field avg7.4

Pages (3 tasks)

Fusion8.5
field avg7.5

Sims (12 tasks)

Fusion8.5
field avg6.7

Visuals (9 tasks)

Fusion8.2
field avg6.7

06 · Outside signals - every row sourced

My bench measures one thing: judged one-shot builds. These outside rows measure other things - kept separate, never blended into the GoldieBench average, every value linked to where it comes from.

From the source guides

MeasureResultSource guide
DRACO deep-research (premium panel)69.0%/fusion-api
DRACO deep-research (budget panel)64.7%/fusion-api
vs solo Fable 5+3.7 pts/fusion-api

07 · Head-to-head records

08 · Methodology + honest limits

Every GoldieBench score is a one-shot, single-file build from an identical prompt - no retries, no hand-fixing - rendered for real, screenshotted, and scored 0-10 by one judge model on one rubric across the whole field. Failures score as failures. What this bench does NOT measure: multi-turn agent work, long-context recall, or API latency - that is what the sourced outside rows are for. Full method: /methodology.

Frequently asked questions

01How does Fusion perform on GoldieBench?

Fusion averages 8.59/10 across 47 scored one-shot build tasks, ranking #1 of 17 ranked frontier models, with 22 task golds, 4 silvers and 8 bronzes. Every score is a real judged build you can open and play on this site.

02What is Fusion best at?

Its strongest scored build is Nordiccrypt at 9.5/10. By category it averages Game 8.8, Page 8.5, Sim 8.5, Visual 8.2 on the bench.

03How much does Fusion cost?

OpenRouter's Fusion API dispatches a single prompt to multiple frontier models and ensembles the answers. Premium panel: Fable 5 + GPT-5.5. Budget panel: cheaper open-weights models. Roughly half the per-token cost of a Fable 5 solo call.

Source ledger

  1. 01Official vendor site: openrouter.ai/openrouter/fusionopenrouter.ai
  2. 02DRACO deep-research (premium panel)agentos.guide

Every model's breakdown

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 4,000+ founders shipping with it every day all live inside the AI Profit Boardroom.

4,000+founders
258documented wins
38countries
$59/momonthly