By ByteBot
Sakana AI shipped Fugu Max on September 11 — a $6-per-million-output-token orchestration engine claiming top scores across six major agentic benchmarks while undercutting Kimi K3 and Sonnet 5 output costs by 40–60%. It is not a new language model. It is a trained coordinator that routes each incoming task across a pool of open-weight and specialized models, then returns a single response. The developer pitch is straightforward: better results, less money, and a one-line API migration.
It’s Not a Model. It’s a Trained Router.
The distinction matters. Fugu Max is built on two ICLR 2026 papers — TRINITY and Conductor — that describe how to train a coordination layer rather than a monolithic model. TRINITY is a roughly 0.6B-parameter coordinator evolved with CMA-ES that assigns Thinker, Worker, and Verifier roles to much larger specialist models behind it. Conductor is a 7B model trained via reinforcement learning to discover natural-language coordination strategies, with the ability to call itself recursively to scale compute at inference time.
The pool of models Fugu Max can route to now includes NVIDIA’s Nemotron family — added via a Sakana-NVIDIA collaboration announced in August — alongside Gemini, DeepSeek, Kimi, Opus, and GPT variants. By dynamically assigning each subtask to the cheapest model capable of handling it, Fugu Max aims to place itself at the efficient frontier of the cost-performance curve rather than simply being the most capable single model.
The Numbers
| Model | Input (per M) | Output (per M) | Context |
|---|---|---|---|
| Fugu Max v1.0 | $2 | $6 | 1M tokens |
| Fugu Ultra v2.0 | $5 | $30 | 1M tokens |
| Kimi K3 (approx) | higher | ~$15 | — |
| Sonnet 5 (approx) | higher | ~$15 | — |
A real-world developer test put Fugu Ultra through the same coding task as Claude Opus 4.8. Fugu Ultra finished in 22 minutes at $7.32. Opus took 79 minutes at $37.85. Opus won on application functionality and design quality; Fugu won on speed and cost. That tradeoff — cheaper and faster, but not necessarily better — shows up consistently in early community feedback.
What the Benchmarks Actually Say
Sakana reports Fugu Max achieves the best overall score across six agentic benchmarks: Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish. It places in the top two on seven of eight benchmarks total. Those are strong numbers. But read the footnotes.
SWEFish is Sakana’s own internal coding benchmark — scoring first on a test you wrote is a yellow flag, not a disqualifier, but it belongs in your thinking. Specific numerical scores are not published, only rankings. The comparisons are bounded to “similarly priced” models, which excludes unconstrained frontier models like Opus 5 or GPT-6 Astra from the table. One independent analysis noted the pricing comparison skipped GPT-5.6 Sol, which sits in a relevant price bracket. None of this invalidates the claim — it means the benchmark story is more carefully constructed than the headline suggests.
How to Switch (It’s One Line)
The API migration is genuinely simple. Fugu Max and Fugu Ultra v2 expose an OpenAI-compatible endpoint — Responses, Chat Completions, and Models APIs — plus an Anthropic-compatible Messages API. If you’re already using the OpenAI Python or JavaScript SDK, swap the base URL and the model string:
from openai import OpenAI
client = OpenAI(
base_url="https://api.sakana.ai/v1",
api_key="YOUR_SAKANA_API_KEY"
)
response = client.chat.completions.create(
model="fugu-max",
messages=[{"role": "user", "content": "your task here"}]
)Both models are also available on OpenRouter if you prefer not to integrate directly. Teams already running the original Fugu can switch to Max or Ultra v2 with the same single-parameter change and no migration work.
Is This Progress or Premium Routing?
The toughest critique comes from Elie Bakouch, a research engineer at Prime Intellect: “This is a closed source orchestrator on top of closed source models. If before you didn’t control the models, now you don’t even control which ones are used or how much.” Sakana frames the swappable pool as supply chain resilience — if one model goes offline or changes its terms, the orchestrator routes around it. That argument has real value for enterprise pipelines exposed to API revocations and sudden service changes. The counter is that Sakana itself becomes the new single point of dependency, and the routing decisions are opaque.
The honest take: Fugu Max’s architectural premise is interesting and the pricing arbitrage is real. The migration is as easy as advertised. Whether the quality tradeoff is acceptable depends entirely on your workload — and the benchmark story deserves a trial in your own environment before you act on it. If you’re running cost-sensitive, high-volume agent pipelines where task completion rates matter more than polished prose, Fugu Max is worth fifteen minutes of testing.
