GoldieBench deep dive
Hermes MoA's full benchmark breakdown.
A panel of frontier models, merged by a chair. The model doesn't matter — the system does. 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.
01 · The headline numbers
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
“Polished WebGL Mandelbrot+Julia explorer with drag-pan, wheel/pinch zoom, double-tap, autopilot flight to curated seahorse targets, orbit-trap filaments/rings, live coordinate readout, and iteration/palette controls — a more complete feature set than Fusion/Opus 4.8 and rivals Kimi's visual depth thanks to smooth coloring and the trap-based detail. Loses a hair to Kimi only because the auto-flight depth is capped by single-precision (no perturbation), but it clearly beats the rest of the field and edges past SOLO Opus.”
“A genuinely polished Three.js wormhole: curved spline tunnel path (not just a straight tube), additive wireframe rings, particles, speed streaks, glow sprites, FOV-warp on boost, and clean pointer-steer + wheel-speed + hold-to-boost controls with HSL color cycling and a vignette. Edges out Fusion and Grok (8.5) on the curving tunnel path and multi-layer depth detail, clearly above SOLO Opus 4.8 (7.5).”
“A polished Neon Breakout with HP bricks, multi-type capsule power-ups (WIDE/SLOW/LIFE), level progression with speed-up, screen shake, flash, lighter-blend particles, starfield, perspective grid, best-score persistence, and full mouse/touch/keyboard control with pause/restart — this edges past Opus 4.8 and Fusion (both 8.5) by combining their juice with deeper systems (HP tiers, power-up variety, level waves) in clean single-file code.”
“The richest aurora build in the field: layered ribbons with composite-lit gradients, vertical light rays, twinkling stars, a lake reflection (mirrored aurora + ripple shimmer), layered mountain silhouettes, occasional meteors, and smooth pointer-steering with color-shift on click. Clearly beats SOLO Opus 4.8 (6.0, no input) and edges past Fugu Ultra (8.0) on detail and interactivity — the reflection and ray work are the differentiators.”
“Spatial-grid boids with clean separation/alignment/cohesion plus predator/beacon pointer modes, scatter burst, live count/speed/vision/separation sliders, and a polished glassmorphic HUD — more interactive and feature-complete than Fugu Ultra (8.5) and well past plain SOLO Opus 4.8 (7.0). The grid optimization, flapping triangle birds, and dual-mode pointer earn it a narrow top spot in this field.”
04 · Where it struggles - quoted, not hidden
Boards that hide the weak rows are brochures. These are Hermes MoA's lowest scored builds, verdicts unedited:
“EYEBALL FAIL: black scene, HUD-only render (title/minimap/bars on void). Playtest said animates but nothing visible.”
“EYEBALL: plays, jungle hill + clouds + clean HUD — but skydiver is a bare red capsule (gate cap). Freefall streaks nice.”
“EYEBALL: cartoon terrain/runway/clouds + tidy HUD — but aircraft is a blob/ball, barely reads as plane (gate cap).”
05 · Category breakdown vs the whole field
Games (23 tasks)
Others (0 tasks)
Pages (3 tasks)
Sims (12 tasks)
Visuals (9 tasks)
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.
Hermes MoA is a custom build inside Julian's Agent OS, not a vendor model - so there is no official benchmark table for it anywhere. The judged GoldieBench scores above are its primary public record.
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 Hermes MoA perform on GoldieBench?
Hermes MoA averages 8.17/10 across 47 scored one-shot build tasks, ranking #2 of 17 ranked frontier models, with 8 task golds, 9 silvers and 4 bronzes. Every score is a real judged build you can open and play on this site.
02What is Hermes MoA best at?
Its strongest scored build is Fractal at 8.7/10. By category it averages Game 8.0, Page 8.0, Sim 8.4, Visual 8.4 on the bench.
03How much does Hermes MoA cost?
Hermes Mixture of Agents dispatches one prompt to a configurable panel of frontier models in parallel, then a named aggregator reads every draft and writes one better final answer. Default panel: Claude Opus 4.8 + GPT-5.5, aggregated by Opus 4.8 — all via the OpenRouter key. Unlike a black-box ensemble, every slot is yours to swap from the Mixture tab in the Agent OS.
Source ledger
- 01Julian's stack page (not a vendor source — this is a custom build, no official vendor exists)aiprofitboardroom.com
- 02Julian's guide: /moa-2-system-beats-modelagentos.guide
Every model's breakdown
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.