Updated October 8, 2026
Fugu Max vs Ultra: Pricing, Tiers, and Who Each Is For
Fugu Max at $2/$6 and Ultra v2 at $5/$30 per million tokens — what each tier buys, who each is for, and the orchestration economics underneath.
The lineup
Sakana sells Fugu in two flavors, and the naming tells you the strategy: pick your optimization target. Fugu Max v1.0 asks "what's the best output per dollar?" Fugu Ultra v2.0 asks "what's the highest capability on hard, multi-step tasks?" Same orchestration engine underneath, different trade-offs. Both are hosted APIs — there are no open weights to self-host — and both speak the OpenAI-compatible API format, so switching your existing code over is mostly a URL change. If the mechanics are fuzzy, start with our explainer on how Fugu's orchestration actually works.
Fugu Max: the value play
Max is priced at $2 per million input tokens and $6 per million output tokens. To put that in perspective, Sakana's own comparison puts its output cost 40 to 60 percent below frontier models like Sonnet 5, GPT 5.6 Terra, and Kimi K3. And it isn't a crippled budget tier: Max claims the best overall score on six benchmarks including Terminal Bench 2.1, GPQAD, and SWEFish, and it leans heavily on open-weight and specialist models — including NVIDIA's Nemotron family, folded in through a collaboration Sakana announced in August 2026.
Who is it for? If your workload is multi-step agentic tasks, heavy tool use, or terminal-style software engineering work — the kind where you're already chaining calls together — Max is the one to test first. One reviewer put it well: for straightforward chat or summarization, a single cheaper model does the same job for a fraction of the cost. Max earns its keep where the orchestration actually has something to orchestrate.
Fugu Ultra v2: the heavyweight
Ultra v2 is priced at $5 per million input tokens, $30 per million output tokens, and $0.50 per million cached input tokens. There's also a premium tier — $10 / $45 / $1.00 — for contexts above 272K tokens. So yes, it's roughly five times the output price of Max. What you buy: the highest capability Sakana can squeeze out of the architecture. On their reported numbers, Ultra v2 takes best or joint-best on five of eight benchmarks, with 48.3 on Chartography (against Opus 5's 27.3) and 74.3 on DeepSWE.
Who is it for? Teams doing complex reasoning, visual reasoning, or hard software engineering benchmarks where combining models measurably helps — and where the cost of a wrong answer dwarfs the cost of the API call. If a single frontier model call already feels expensive, Ultra v2 will feel worse; if you're currently chaining three frontier calls to get a reliable answer, one orchestrated call may come out ahead.
The economics underneath
The pricing war has quietly moved. It used to be about individual model inference costs — who trains the cheapest giant. Fugu shifts the battlefield to orchestration economics: the insight that sending a simple lookup to a multi-trillion-parameter model wastes money, and a system that picks the leanest machinery capable of solving each part can beat a single flagship on both price and quality. That's the decoupling Sakana is selling — performance without proprietary lock-in.
Practical notes before you sign up
A few things worth knowing. First, Fugu is hosted-only: no weights to download, no self-hosting. Second, Sakana has held back availability in the EU and EEA pending GDPR compliance work, so it isn't a safe pick for a production stack serving European users right now. Third, all the benchmark claims are Sakana's own — no independent evaluation is cited in the announcements, so treat the numbers as directional until third parties weigh in. And finally, remember the bigger context: the reason orchestration-over-open-models is strategically interesting goes beyond your API bill. Here's why it matters for lock-in, resilience, and export controls.