
Real head-to-head · same prompt, one shot
MiniMax M3 vs Fugu Ultra
1M-context frontier model at $0.30/M tokens — cheapest big-context model on the bench. vs Sakana's multi-agent answer to Fusion — frontier ensemble without single-vendor risk.
Head-to-head verdict: Fugu Ultra wins 3–1 with 1 tie.
What I tested — same prompt, two models
I run the same fixed prompt set through every new model the day it drops — same string, one shot, single HTML file out — and I score the result 0–10 on whether it ran, how close it hit the brief, and how good it looked. Below is what came out when I gave the exact same prompts to MiniMax M3 and Fugu Ultra, side by side, on 5 shared tasks inside the Agent Operating System.
Both models were given identical prompts inside the Agent Operating System — no help, no iteration, no "best of N" tricks. I run each prompt once, save the HTML file the model produces, and score it 0–10 on whether it ran, how close it hit the brief, and how good it looked. The scoring is mine. The verdicts below are pulled from my source comparison guides at agentos.guide where I publish every score and the reasoning behind it.
MiniMax M3 · Bench prompts dispatched via OpenRouter. Scored by Claude judge against the same 42 prompts every other model ran.
Fugu Ultra · Dispatched from Agent OS as the panel-ensemble alternative to OpenRouter Fusion. Bench scored by Claude judge against the same 42 prompts as every other model.
Side-by-side on 42 shared tasks
Click any cell to play that model's actual one-shot attempt. Medals are derived from my 0–10 scores per task (highest = 🥇, second = 🥈, third = 🥉).
Task ↓
MiniMax M3
Fugu Ultra
Game
Page
Sim
Sim
Visual
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Page
— not attempted —
Where MiniMax M3 beat Fugu Ultra
The tasks where I gave MiniMax M3 a higher 0–10 score on the same prompt — with the actual commentary from my source guides.
Voxel
Visual
MiniMax M3 8.0
·
Fugu Ultra 3.5
(+4.5)
What I saw: 29KB Temple-Run-style voxel runner on three.js — lane switching, jump + slide, coins.
Where Fugu Ultra beat MiniMax M3
The tasks where I gave Fugu Ultra a higher 0–10 score on the same prompt — with the actual commentary from my source guides.
Raycaster
Game
Fugu Ultra 8.5
·
MiniMax M3 7.0
(+1.5)
What I saw: 26KB canvas raycaster with WASD + mouse-look + distance fog + weapon bob. Clean implementation, comparable to Fusion's 17KB on the same prompt. ~$0.35 per call — roughly 1/4 the cost of Fusion.
Galaxy
Sim
Fugu Ultra 8.5
·
MiniMax M3 7.5
(+1.0)
What I saw: 26KB three.js spiral galaxy with drag-to-orbit + dust lanes + bloom. Comparable visual quality to Fusion's 14KB attempt with more polish on the camera UI. ~$0.24 per call.
Landing
Page
Fugu Ultra 9.0
·
MiniMax M3 8.0
(+1.0)
· winner · denser build
What I saw: 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.
Strengths & weaknesses I logged
MiniMax M3
Strengths
- 1M token context — full repo / full deep-research corpus fits in one call
- $0.30/M input is roughly 1/30th of Opus 4.8 — built for high-volume agent loops
- Solid one-shot HTML output — clean structure on game and visual prompts
Trade-offs
- Less polished than Fusion's panel-ensembled output on the toughest deep builds
- Newer model — less community calibration vs Fable 5 / Opus 4.8
Fugu Ultra
Strengths
- SWE Bench Pro 73.7 · GPQA-D 95.5 · MRCRv2 93.6 — Sakana's published frontier-tier benchmark scores
- Vendor-agnostic ensemble — opt out of specific providers for compliance / export-control
- OpenAI-compatible API at api.sakana.ai — drop-in for existing tooling
Trade-offs
- Panel orchestration adds latency — even a 'pong' burns ~2k orchestration tokens
- Newer than Fusion; less community calibration on long-tail prompts
Pricing & context — the spec sheet
| Spec | MiniMax M3 | Fugu Ultra |
|---|---|---|
| Vendor | MiniMax | Sakana AI |
| Context window | 1,048,576-token context — matches GLM-5.2 and Fable 5 | 272,000 tokens with the standard rate. Calls exceeding 272K context are billed at the higher 'long-context' rates. |
| Price | $0.30 / 1M input tokens, $1.50 / 1M output | $5 / 1M input · $30 / 1M output (Fugu Ultra) |
| Pricing detail | MiniMax M3 is the cheapest 1M-context frontier model on the bench — roughly 1/200th the per-call cost of OpenRouter Fusion and 1/30th of Claude Opus 4.8. Designed for high-volume agent workloads where context length matters but per-call budget is tight. | 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. |
| Release | 2026-06-18 | 2026-06-15 |
| Bench coverage | 42/42 scored · avg 7.96/10 | 5/5 scored · avg 7.60/10 |
The verdict — which should you pick?
Across 5 scored shared tasks, the averages are essentially tied — MiniMax M3 7.80 vs Fugu Ultra 7.60. This isn't the comparison where one wins; it's the comparison where you pick based on context, pricing, and what you're actually trying to ship.
If you only run one of these inside your stack, the head-to-head average above is the call. If you can run both, my honest play is to wire MiniMax M3 and Fugu Ultra both into the Agent Operating System and dispatch each from the kanban by task type — high-volume agent workflows where per-call cost dominates → MiniMax M3, teams that want fusion-class quality but need a different vendor risk profile → Fugu Ultra. That's the same setup I run for the 3,600+ founders inside the AI Profit Boardroom.
FAQ — MiniMax M3 vs Fugu Ultra
Which is better, MiniMax M3 or Fugu Ultra?
On Goldie Bench, MiniMax M3 averages 7.80/10 across the shared tasks, with 12 gold, 11 silver, 8 bronze overall. Fugu Ultra averages 7.60/10, with 3 gold, 1 silver, 0 bronze. Fugu Ultra wins the head-to-head 3–1.
How much does MiniMax M3 cost vs Fugu Ultra?
MiniMax M3: MiniMax M3 is the cheapest 1M-context frontier model on the bench — roughly 1/200th the per-call cost of OpenRouter Fusion and 1/30th of Claude Opus 4.8. Designed for high-volume agent workloads where context length matters but per-call budget is tight. Fugu Ultra: 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.
What's the context window for MiniMax M3 vs Fugu Ultra?
MiniMax M3 has a 1,048,576-token context — matches GLM-5.2 and Fable 5 context window. Fugu Ultra has a 272,000 tokens with the standard rate. Calls exceeding 272K context are billed at the higher 'long-context' rates. context window.
When should I pick MiniMax M3 over Fugu Ultra?
Pick MiniMax M3 for: High-volume agent workflows where per-call cost dominates; 1M-context tasks (whole-repo refactors, deep-research synthesis); Drop-in cheaper alternative to GLM-5.2 with comparable 1M context. The trade-off is the weaknesses we logged on the bench: Less polished than Fusion's panel-ensembled output on the toughest deep builds; Newer model — less community calibration vs Fable 5 / Opus 4.8.
When should I pick Fugu Ultra over MiniMax M3?
Pick Fugu Ultra for: Teams that want Fusion-class quality but need a different vendor risk profile; Operators avoiding export-controlled providers (Sakana emphasises this in their pitch); Deep-research workflows where ensemble verdicts beat single-model answers. The trade-off is the weaknesses we logged on the bench: Panel orchestration adds latency — even a 'pong' burns ~2k orchestration tokens; Newer than Fusion; less community calibration on long-tail prompts.
How does Goldie Bench score MiniMax M3 vs Fugu Ultra?
Every demo on this page was built by Julian Goldie inside the Agent Operating System — same fixed prompt for both models, one shot, single HTML file out. Each result gets a 0–10 score on whether it ran, how close it hit the brief, and how good it looked. The highest score on each task gets gold; second gets silver; third gets bronze. See methodology for full provenance.
Related comparisons
Other head-to-heads using the same scoring system:
MiniMax M3 vs Opus 4.8 Fugu Ultra vs Opus 4.8 MiniMax M3 vs GLM-5.2 Fugu Ultra vs GLM-5.2 MiniMax M3 vs Grok Fugu Ultra vs Grok MiniMax M3 vs Fusion Fugu Ultra vs FusionFull model pages: MiniMax M3 · Fugu Ultra · back to the leaderboard
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 3,600+ founders shipping with it every day all live inside the AI Profit Boardroom.
3,600+founders
258documented wins
38countries
$100k+/mocommunity MRR



























