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

GoldieBench deep dive

Fugu Ultra's full benchmark breakdown.

Sakana's multi-agent answer to Fusion — frontier ensemble without single-vendor risk. 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
42
Reading time
9 min
External sources
3

01 · The headline numbers

GoldieBench average
7.94/10
42 scored one-shot tasks
Board rank
#8
of 17 ranked frontier models
Task medals
5🥇 2🥈 7🥉
outright wins on shared briefs
Context
272,000 tokens with the standard rate. Calls exceeding 272K context are billed at the higher 'long-context' rates.
$5 / 1M input · $30 / 1M output (Fugu Ultra)

02 · Every benchmark, bar by bar

All 42 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

“Ultra v2 (gap-fill) — click-to-launch fireworks. Smoke-test PASS with 26.3% pixel diff — among the most reactive builds.”

- the judge's verdict, unedited

“Ultra v2 — juicy browser game. Smoke-test PASS with 24.7% pixel diff — one of the most reactive builds on the bench.”

- the judge's verdict, unedited

“Sakana Fugu Ultra shipped a 32KB Apple-keynote landing — bigger than Fusion's 20KB attempt at the same prompt. Animated mesh gradient, multi-section, polished. $0.32 vs Fusion's $1.30 for the same output — 4× cheaper, denser result.”

- the judge's verdict, unedited

“Ultra v2 (gap-fill) — 61.5KB Nordic dungeon crawler with bloom + boss room. Smoke-test PASS with 22.8% pixel diff — highly reactive.”

- the judge's verdict, unedited

“Fugu Ultra v2 rebuild — 55.7KB solar system, the densest solar attempt on the bench. Full </html>, animation loop, Saturn rings, drag-to-orbit + scroll-to-zoom. Smoke-test PASS (3.6% pixel diff after drag, zero console errors). The panel ensemble produces a markedly richer build than Mini's 12KB single-model version.”

- the judge's verdict, unedited

04 · Where it struggles - quoted, not hidden

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

“Ultra v2 — billiards. Smoke-test MAYBE (0.0% diff) — cue needs click-drag aim that generic input didn't replicate.”

- the judge's verdict, unedited

“Ultra v2 — arcade shooter. Smoke-test MAYBE (0.0% diff) — may need a specific start interaction; flagged for manual verification.”

- the judge's verdict, unedited

“Ultra v2 — fluid sim. Smoke-test MAYBE (0.0% diff) — needs click-drag injection which generic input didn't replicate.”

- the judge's verdict, unedited

“Ultra v2 — wormhole tunnel. Smoke-test MAYBE (0.3% diff) — animates slowly; hold-space accel didn't show big change in the short test.”

- the judge's verdict, unedited

05 · Category breakdown vs the whole field

Games (19 tasks)

Fugu Ultra7.7
field avg7.1

Others (0 tasks)

field avg7.4

Pages (2 tasks)

Fugu Ultra8.0
field avg7.5

Sims (12 tasks)

Fugu Ultra8.1
field avg6.7

Visuals (9 tasks)

Fugu Ultra8.2
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
SWE Bench Pro73.7/sakana.ai/fugu
GPQA-Diamond95.5/sakana.ai/fugu
MRCRv293.6/sakana.ai/fugu

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 Fugu Ultra perform on GoldieBench?

Fugu Ultra averages 7.94/10 across 42 scored one-shot build tasks, ranking #8 of 17 ranked frontier models, with 5 task golds, 2 silvers and 7 bronzes. Every score is a real judged build you can open and play on this site.

02What is Fugu Ultra best at?

Its strongest scored build is Fireworks at 9.0/10. By category it averages Game 7.7, Page 8.0, Sim 8.1, Visual 8.2 on the bench.

03How much does Fugu Ultra cost?

Sakana's multi-agent orchestration: a single API call internally dispatches to multiple frontier models and synthesises the answer. Subscription plans run $20-$200/mo (Standard / Pro / Max); PAYG is $5/M input + $30/M output for Fugu Ultra. Direct competitor to OpenRouter Fusion's panel approach.

Source ledger

  1. 01Official vendor site: sakana.aisakana.ai
  2. 02SWE Bench Proagentos.guide
  3. 03Julian's guide: /sakana-fugu-vs-fusionagentos.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