Fugu Ultra: What It Is, What It Costs, When It’s Worth It

Fugu Ultra is what Sakana AI ships when the answer matters more than the invoice. Where base Fugu balances performance against latency, Ultra draws on a deeper pool of expert agents and routes between one and three of them per problem — a system built for the questions you’d otherwise hand to your most senior person and a reviewer. This guide covers how it differs from base Fugu, what its reasoning-effort levels really do, what it costs, and when it’s worth it. Facts verified against Sakana’s official pages on August 19, 2026.

What Fugu Ultra is

Like every member of the family, Ultra is a trained orchestrator: one API-facing model that reads your task, assembles a team from a pool of frontier models, coordinates their work, and returns a single answer. What makes it “Ultra” is scale and depth on both sides of that process — a larger pool of expert agents to draw from, and the willingness to spend more internal work per problem. Sakana positions it for “hard, high-stakes tasks”: the technical report describes Fugu Ultra as the model that prioritizes answer quality on the hardest problems.

Two spec anchors: a 1,000,000-token context window (with a pricing tier change above 272K tokens of context), and support for image input alongside text. Since an August 2026 update, Ultra is offered as explicit pinned versions — selecting one means a later Ultra release never changes the model under you without your say-so.

How Ultra differs from base Fugu

	Fugu	Fugu Ultra
Routing	Picks the best model per task from the supported pool	Routes between 1–3 expert agents per problem, from a deeper pool
Optimized for	Balanced performance and latency, everyday work	Answer quality on hard, high-stakes problems
Billing	The top-tier underlying model’s standard rate, never stacked	Fixed per-token rates (plus billed orchestration usage)
Effort levels	high, xhigh	high, xhigh, max (current version only)
Cost predictability	Variable — depends on routing	Fixed rates, variable orchestration overhead

The “1–3 agents” line is the architectural heart of it. Base Fugu’s job is mostly to pick the right specialist; Ultra’s job is to decide whether a problem deserves a team — and if so, to compose one, brief each member, and verify the result. That extra machinery is also why Ultra has the billing property that surprises newcomers: its internal agents’ work is metered as orchestration tokens, billed at full standard rates on top of your visible traffic.

Reasoning effort levels, explained

Ultra exposes a dial most models don’t: high, xhigh, and max. It controls how much internal reasoning and coordination a request gets — more effort means better odds on hard problems, more tokens, higher cost, and longer turns.

• high — the default, balancing speed and performance. Sakana’s own coding-agent integration ships with this as its starting setting, which says a lot about where the sweet spot sits.
• xhigh — deep reasoning for complex problems.
• max — a distinct maximum level for the hardest problems, and an Ultra exclusive: per Sakana’s documentation, only the current Ultra version treats max as its own level. Everywhere else in the family, max is accepted for compatibility but behaves as xhigh. (Sakana’s own pages briefly disagreed on this detail; the get-started documentation and the machine-readable model catalog — which we follow — both confirm it as of August 19, 2026.)

Practical guidance: start at high, and escalate per task, not per project. Effort is a request-level setting, so a pipeline can run 95% of its traffic at high and reserve max for the handful of problems that earn it.

What Fugu Ultra costs

Ultra’s pay-as-you-go rates per million tokens (as of August 2026, applying to both current versions):

Token type	Standard	Context > 272K
Input	$5	$10
Output	$30	$45
Cached input	$0.50	$1.00

Three cost facts to carry with the table. Orchestration usage is billed on top of visible tokens at these same rates. Output-length caps don’t cap request costs — the orchestrator keeps its own budget. And every subscription plan ($20–$200/month) includes Ultra, which for hands-on use converts all of this variability into a flat fee. Model a real workload in the cost calculator — it has an orchestration-overhead slider and the long-context toggle for exactly this.

What people actually use Ultra for

Sakana’s published use cases for Ultra cluster where verification is worth paying for: performance in competitions, reproducing results from research papers, and patent and literature review — all reported by Sakana itself. The pattern across them: long, multi-step problems with a checkable right answer, where a Thinker/Verifier team structure earns its overhead.

On benchmarks, Sakana’s June 2026 evaluation reports Ultra at 73.7 on SWE Bench Pro (base Fugu: 59.0) — the widest gap in the family’s favor — plus 82.1 on TerminalBench 2.1 and 95.5 on GPQA-Diamond, against baselines drawn from leading frontier models. Treat these as the vendor’s own measurements, dated June 2026; the full comparison is on Sakana’s site.

When to choose Ultra — and when not to

Choose Ultra when the cost of a wrong answer dwarfs the cost of the request: architecture decisions, subtle bugs, research-grade analysis, high-stakes review. Stay on base Fugu for everyday coding and reasoning — Sakana’s own benchmark table shows base Fugu matching Ultra on some tests and even leading on a few, at a fraction of the cost and latency. The honest heuristic: if you wouldn’t convene a meeting of your best people over the question, don’t convene Ultra’s agents either.

The full decision framework, with scenarios and the cost math side by side, is in Fugu vs Fugu Ultra.

Go deeper: how Ultra’s billing really works, the base-vs-Ultra decision guide, or hands-on in the playground.