What is Sakana Fugu and how is it different from a standard LLM?

Sakana Fugu is not a single trained model. It is a learned orchestration system that routes each incoming task to the most capable and cost-efficient model in a pool of open-weight and specialized models, then combines their outputs behind one OpenAI-compatible API endpoint. Users send one request; Fugu decides which underlying models to involve.

How much do Fugu Max and Fugu Ultra v2 cost?

Fugu Max is priced at $2 per million input tokens and $6 per million output tokens, with a flat rate regardless of context length. Fugu Ultra v2 is $5 per million input tokens and $30 per million output tokens, with a higher rate above 272K tokens ($10/$45). Sakana says Fugu Max output pricing is 40 to 60 percent below Claude Sonnet 5, GPT-5.6 Terra, and Kimi K3.

Is Sakana Fugu available to enterprise teams in Europe?

No. Sakana has stated that Fugu Max and Fugu Ultra v2 are not available in the EU or the EEA while the company works on GDPR compliance. Teams outside Europe can access both through Sakana's hosted API or through OpenRouter.

What benchmarks did Fugu Ultra v2 top, and are they independently verified?

Sakana reports that Fugu Ultra v2 scores 48.3 on Chartography (vs 27.3 for Claude Opus 5 and 29.5 for Claude Fable 5) and 74.3 on DeepSWE. It achieves best or joint-best on 5 of 8 reported benchmarks. As of launch date, these are Sakana's own evaluation results and have not been independently reproduced by a third party. SWEFish is an internal Sakana benchmark.

Can Fugu Max or Ultra v2 be self-hosted?

No. Model weights are not released. Both variants are available only through Sakana's hosted API. Users who need on-premise or self-hosted deployment cannot use either product in its current form.

Sakana Fugu Max and Ultra v2: Orchestration Beats Frontier

Sakana AI released Fugu Max and Fugu Ultra v2 on September 11, 2026, extending its bet that a learned orchestration layer routing queries across a pool of models can outperform single-model architectures on both cost and capability. Fugu Max targets the cost-performance frontier at $2 per million input tokens. Fugu Ultra v2 targets peak output quality, reporting benchmark scores above Claude Opus 5 and Claude Fable 5 on visual reasoning and real-world software engineering without using either in its agent pool.

Both models are available today through Sakana’s OpenAI-compatible API with a single-line parameter switch for existing Fugu customers. Neither is available in the EU or EEA.

What Fugu Actually Is

The central claim of the Fugu architecture is that orchestration replaces scale. Rather than training one enormous model on a large compute budget, Sakana trains a routing model: a language model whose job is to decompose an incoming task, select the right combination of underlying models from a fixed pool, coordinate their outputs, and return a single coherent response to the caller.

The pool includes open-weight models (including NVIDIA’s Nemotron family, integrated through Sakana’s partnership with NVIDIA) and specialized models tuned for domains like coding and security. The routing logic and pool composition are proprietary and are not exposed to API callers.

From the caller’s perspective, Fugu looks identical to any frontier model: one endpoint, one API key, one stream of output tokens. From Sakana’s perspective, a single request may touch one model or several, and billing is based on the top-tier model involved in any given request rather than stacking each model’s fee.

The product has moved fast since its April 2026 beta. Fugu reached general availability in June alongside Fugu Ultra v1. July added Fugu-Cyber, a variant tuned for security reasoning, and a Claude Code interface. August added Sakana Chat (consumer) and started integrating NVIDIA Nemotron. September brings Fugu Max and Fugu Ultra v2.

Fugu Max: Cost-Performance at $2/M Tokens

Fugu Max is built to beat the economics of running frontier-class agents at production scale. Sakana prices it at $2 per million input tokens and $6 per million output tokens, a flat rate with no long-context surcharge. Cached input is $0.25 per million tokens.

Sakana says this output rate is 40 to 60 percent below Claude Sonnet 5 ($15/M output), GPT-5.6 Terra, and Kimi K3. That gap becomes significant when you are running agents at scale: a team processing a million output tokens per day would spend $6 with Fugu Max versus $15 with Sonnet 5, a difference of $2,700 per month before any volume adjustments.

On benchmarks, Sakana reports Fugu Max achieves the best overall score on six of ten tested metrics: Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench, and SWEFish. It also claims to extend the cost-performance Pareto frontier (matching or exceeding quality at lower cost) on seven of ten benchmarks evaluated.

The practical use case is high-volume agent work: code review agents running across large repositories, document extraction pipelines processing thousands of files per day, or structured data workflows where per-token cost compounds into meaningful monthly spend. Fugu Max routes each subtask to the cheapest model that can solve it; the caller sees one response at one price.

Fugu Ultra v2: Capability Without Closed Dependencies

Fugu Ultra v2 is aimed at the hardest multi-step reasoning, autonomous research, and full-stack software engineering tasks. Its architecture is the same routing system; the tuning pushes toward maximum quality at the cost of latency and per-token spend.

The benchmark Sakana emphasizes most is Chartography, which tests visual reasoning and data interpretation from charts, diagrams, and structured data. Fugu Ultra v2 scores 48.3 against Claude Opus 5 at 27.3 and Claude Fable 5 at 29.5. On DeepSWE, a benchmark designed around real-world software engineering rather than isolated code generation, Fugu Ultra v2 scores 74.3, which Sakana says beats models priced three to five times higher per token.

The detail Sakana calls out explicitly: Fugu Ultra v2 does not include Claude Fable 5, Claude Fable 5.1, or GPT-6 Astra in its agent pool. The architecture cannot use those models because they are not publicly accessible via API. The claim is that orchestrating open-weight and mid-tier models correctly is enough to reach or exceed the output quality of those closed flagships.

That claim has not been independently verified as of September 13. All benchmark numbers are Sakana’s own evaluations. SWEFish is an internal Sakana benchmark. Third-party reproduction of the Chartography and DeepSWE results would substantially strengthen the case.

Model	Chartography	DeepSWE	Input $/M	Output $/M
Fugu Ultra v2	48.3	74.3	$5	$30
Claude Opus 5	27.3	N/A	$15	$75
Claude Fable 5	29.5	N/A	$3	$15
Fugu Max	N/A	N/A	$2	$6

Benchmark scores from Sakana AI’s September 11, 2026 release. Claude pricing from Anthropic’s API pricing page. N/A indicates the metric was not reported in Sakana’s comparison.

The NVIDIA Nemotron Integration

The Fugu Max pool now includes NVIDIA’s Nemotron family of open-weight models, integrated through the partnership Sakana disclosed in August 2026. This is worth noting beyond the marketing: NVIDIA’s strategy has been to ship open models (Nemotron, Llama-based variants, domain-specific checkpoints) that make its GPU infrastructure more valuable by increasing the variety of capable models developers can run on that infrastructure.

Fugu Max turning NVIDIA’s open models into part of a production-grade orchestration layer validates that strategy. Enterprises using Sakana get Nemotron capabilities without managing those model deployments themselves; NVIDIA gets its models deployed at scale through a third-party product. The arrangement is not disclosed in technical detail, but it establishes the pattern: open model providers benefit from orchestration layers that make their models easy to consume.

What This Means for Enterprise AI Teams

The practical question for an enterprise AI team is not whether Fugu’s benchmark numbers hold up at their stated values. It is whether the architectural proposition changes anything about how you should think about model procurement.

For the past two years, the implicit model has been: pick one frontier provider, standardize your stack on it, and accept that you are dependent on that provider’s pricing, availability, and capability roadmap. The alternative (building your own multi-agent harness) is expensive in engineering time and difficult to maintain as underlying models change.

Sakana’s bet is that a managed orchestration layer gives you the routing benefits without the maintenance cost, and that the routing layer can stay vendor-agnostic even as the pool of underlying models changes. If a model in the pool changes price or availability, Sakana updates the routing; the caller sees no change.

The enterprise AI governance implications are real. Orchestration layers that choose which models see your data, on your behalf, without exposing that selection to you, create audit and compliance complexity. You cannot log which model processed a given request if Sakana’s routing is opaque. For regulated industries, that opacity may be disqualifying regardless of cost savings. Sakana allows users to exclude specific agents from the pool for compliance reasons, but confirming that exclusions are enforced requires trusting Sakana’s assurances rather than verifying them directly.

For teams outside regulated industries with meaningful monthly AI spend and latency tolerance for orchestration overhead, Fugu Max deserves a direct cost-per-completed-task comparison against your current model. Token price per million is the wrong metric; the right one is cost per finished task including all orchestration tokens consumed.

Enterprise AI-native transformation increasingly depends on cost-predictable inference. Fugu Max’s flat pricing structure (no long-context surcharge on Max, known overage rates on Ultra v2) is a planning advantage over models with tiered pricing that can double at a context threshold.

Limitations Worth Tracking

Several things about this launch are not yet resolved:

The benchmark numbers are Sakana’s own. SWEFish is internal. The Chartography and DeepSWE comparisons use Sakana’s evaluation harness, and the comparison scores for Opus 5 and Fable 5 are not clearly attributed to provider-published evaluations in the launch materials. Independent reproduction is the next test.

Output reproducibility is constrained. Routing can select different models for identical prompts, which means output style and reasoning traces may vary across runs. Teams that need deterministic outputs for compliance or debugging purposes should plan around this.

EU/EEA teams are excluded entirely while Sakana completes GDPR compliance work. No timeline has been given.

The supply-chain independence argument has a counterpoint. Depending on a single managed orchestration provider (whose pool and routing logic are opaque) is still a single vendor dependency. The dependency has shifted from model to orchestrator.

Both models launch through the hosted API at OpenRouter alongside Sakana’s own endpoint. No weights are available for self-hosting.

Sources