Fugu Ultra: Sakana AI Model, Pricing, Benchmarks & Use Cases

Last updated: June 23, 2026

Fugu Ultra is Sakana AI’s quality-first version of Sakana Fugu, a multi-agent model orchestration system exposed through one OpenAI-compatible API. If standard Fugu is the lower-latency default, Fugu Ultra is the variant Sakana positions for harder multi-step tasks where answer depth matters more than response time.

Quick answer: Fugu Ultra is not just a bigger single chatbot. It is part of Sakana Fugu, a learned orchestration system that can coordinate a pool of expert AI agents. Sakana lists fugu-ultra-20260615 pricing at $5 per 1M input tokens, $30 per 1M output tokens, and $0.50 per 1M cached input tokens, with higher rates for context above 272K tokens. Sakana’s published benchmark table shows strong results, but those are vendor-published results and should be validated on your own workload.

For the broader launch and Claude Fable 5 comparison, read the main Sakana AI Fugu review. This page focuses only on the Fugu Ultra query cluster: model type, pricing, benchmarks, use cases, and evaluation checklist.

What is Fugu Ultra?

Fugu Ultra is the performance-oriented variant of Sakana Fugu. Sakana describes Fugu as a multi-agent system delivered as one model: you call one API, while Fugu handles model selection, routing, coordination, and synthesis behind the scenes.

The difference is operating mode:

Sakana’s Fugu product page says Fugu and Fugu Ultra are both available through one OpenAI-compatible API. That matters for developers because you can test the two modes without rebuilding your whole integration.

Fugu Ultra vs Fugu

The simplest way to choose is to start with the task shape.

Use Fugu first when:

• the user is waiting interactively;
• latency matters;
• the task is routine coding, editing, or summarization;
• you need configurable agent opt-outs for data, privacy, or compliance;
• a cheaper or simpler model can solve the task reliably.

Evaluate Fugu Ultra when:

• the task is multi-step and failure is expensive;
• the work benefits from planning, verification, and synthesis;
• the prompt contains a paper, codebase, dataset, patent set, or long document set;
• answer depth matters more than immediate response time;
• you are comparing against premium models such as Claude Fable 5, Opus 4.8, GPT-5.5, or Gemini 3.1 Pro.

Sakana’s own examples for Fugu Ultra include Kaggle competitions, paper reproduction, cybersecurity analysis, and literature or patent investigations. Treat those as starting points for evaluation rather than fixed expectations.

Fugu Ultra pricing

Sakana lists fixed pay-as-you-go pricing for fugu-ultra-20260615:

Sakana also lists monthly subscription tiers that include access to both Fugu and Fugu Ultra:

The important evaluation metric is cost per completed task, not only price per token. Fugu Ultra may use hidden orchestration work behind the final answer. Sakana’s pricing page explains that orchestration tokens can count toward usage. For a fair test, log input, output, cached-input, orchestration usage, latency, retry rate, and human review time.

Fugu Ultra benchmark results

Sakana’s published Fugu table reports the following Fugu Ultra scores:

These numbers are useful, but they should be read with the caveat Sakana itself gives: baseline scores are provider-reported, and the Fugu results are from Sakana’s June 2026 evaluation. Teams should not make production routing decisions from the table alone.

A better internal test is to choose 20-50 representative tasks and compare:

• success rate;
• answer quality after human review;
• cost per successful task;
• latency;
• retry count;
• policy or compliance failures;
• whether the final answer is easier to audit.

Fugu Ultra vs Claude Fable 5

Fugu Ultra and Claude Fable 5 are often compared because both target difficult long-horizon work. The comparison is useful, but the products are structurally different.

The practical question is not “which one wins?” It is: which option fits your workload, compliance constraints, latency tolerance, and provider-risk model? If Fable access is unavailable or your team wants a provider-flexible orchestration layer, Fugu Ultra may be worth evaluating. If you need a single known model identity and Anthropic access is available, Fable 5 may be simpler to govern.

Is Fugu Ultra open source?

No — not in the sense of “the full hosted model orchestration service is open source.” Sakana has a public SakanaAI/fugu GitHub repository, and that repository includes setup docs, configuration materials, the README, assets, and the technical report. But the hosted Fugu and Fugu Ultra service is an API product.

A procurement-friendly distinction:

• public GitHub repo: yes;
• public technical report: yes;
• open-source hosted orchestration service: no;
• exact worker routing visible to users: no;
• API product with documented pricing: yes.

When should you test Fugu Ultra?

Test Fugu Ultra when the work is hard enough that orchestration may matter. Good candidates include:

• long codebase review;
• migration planning;
• paper reproduction;
• patent and literature analysis;
• Kaggle-style data science;
• research synthesis;
• scoped security review with evidence requirements;
• complex tasks where a single-model answer often needs multiple retries.

Do not start with Fugu Ultra for every task. For short writing, simple summarization, quick Q&A, or basic code snippets, a lower-latency and lower-cost model may be enough.

Sources checked

• Sakana Fugu product page
• Sakana Fugu launch announcement
• SakanaAI/fugu GitHub repository
• Claude Fable 5 page
• Anthropic statement on Fable 5 and Mythos 5 access

FAQ