Published: 2026-06-23

Sakana Fugu Ultra vs Claude Opus 4.8: 38-Task Battle Test

Chapters / key moments (click to jump — plays here on the page)

Nate Herk takes Sakana's viral Fugu Ultra — a single API that orchestrates frontier models (Opus, GPT, Gemini) like a multi-agent router — and runs it head-to-head against Claude Opus 4.8 across 38 tasks. The result: 36 ties, with Fugu roughly 4.5× slower and 5× more expensive, largely because Opus is one of the very models Fugu delegates to.

Source video

"I Battle Tested Sakana Fugu's Fable Killer" by Nate Herk — Watch on YouTube →

Key Takeaways

Fugu is not a new LLM. It's a small "manager" model that breaks a task down and routes sub-tasks to frontier models (Opus, GPT-5.5, Gemini, and others), then has another model merge the results — a multi-agent system delivered as one API.
It runs inside Claude Code via a markdown config file plus an API key. Notably, the context window stays near zero through a long session because responses are routed through Fugu's server rather than filling Claude Code's own context.
The scoreboard: across 38 AI-generated, Codex-graded, mostly pass/fail tasks (puzzles, traps, specs, heavy algorithms), 36 ended in ties and Opus won 2. Fugu never clearly won — unsurprising, since Opus 4.8 is one of the models Fugu itself selects from.
Cost and speed are the story: Fugu's runs took 357 minutes total vs Opus's 80 minutes, and cost ~$50 vs ~$10 — about 4.5× slower and 5× pricier. Easy tasks Opus answered in ~6 seconds took Fugu several minutes.
The pattern isn't new. It's the same orchestration you already do pairing Claude Code sub-agents, or running Codex and Claude Code on one codebase — Fugu just automates the delegation. It differs from OpenRouter's Fusion API, which fans the same prompt to three models and judges/merges rather than splitting the task.
Honest takeaway: impressive benchmarks, but for knowledge work the cost and latency aren't worth it over a Claude Code or Codex subscription. The real value is for heavy, multi-team software development — and the broader skill of optimizing which model does which task is only getting more important.

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

What Fugu is
Sakana's Fugu is a multi-agent system presented as one model behind a single API: a small manager model breaks a task down and routes parts to other frontier models (the speaker names Opus, GPT and Gemini) and then combines the answers. He contrasts it with OpenRouter's Fusion API, which sends one prompt to several models and merges the answers with a judge.
How he tested it
He had Codex write 38 pass/fail style tasks in four waves (puzzles, traps, specs, heavy algorithms), ran the same prompts through Fugu Ultra and Claude Opus 4.8 on API billing, and had Codex grade both.
Results he reported
36 of 38 tasks tied; Opus won the other two. Fugu Ultra took about 357 minutes in total against about 80 for Opus (roughly 4.5x slower) and cost about $50 against about $10 (roughly 5x). Easy tasks Opus answered in seconds took Fugu minutes.
Using it inside Claude Code
He ran Fugu Ultra inside Claude Code using a markdown file (shared in his community) plus an API key; it isn't a simple endpoint swap. The Claude Code context meter stayed near zero in long chats because of how requests route to Fugu's server. Plans: pay-as-you-go API billing or a subscription (he used the $200 per month plan after upgrading from $100).

Gotchas

His own verdict: he won't use it for knowledge work; it may suit heavy software teams that want a GPT reviewer and a Claude planner behind one API.
The test suite was AI-written and AI-graded, and Opus 4.8 is one of the models Fugu can route to, so treat the comparison as indicative.
He notes the vendor's own claim that Fugu outperforms Fable and Mythos preview on some benchmarks is about orchestration, not a better base model.