Sakana Fugu vs Fugu Ultra: Full Comparison Guide

When Sakana launched Fugu on June 22, 2026, they did not ship one model. They shipped two: Fugu and Fugu Ultra. Same API, same architecture, two different points on the speed-versus-quality curve.

The confusion most people have is not about what Fugu is. It is about which variant they should actually be using. Default to Fugu and you may be leaving real quality on the table for hard problems. Default to Fugu Ultra and you may be paying double the time and tokens for a tier of capability your workload does not need.

This guide breaks down exactly how the two variants differ, what each one is genuinely better at, and how to route between them so you are not over-paying or under-performing.

The One-Line Difference

Fugu coordinates a smaller pool of agents and prioritizes low latency. Good for everyday work where speed matters and quality is "good enough."

Fugu Ultra coordinates a deeper pool of expert agents and prioritizes maximum quality. Good for hard, multi-step problems where getting the right answer matters more than getting a fast answer.

Both run on the same orchestration architecture, built on Sakana's TRINITY and Conductor research papers from ICLR 2026. Both expose the same OpenAI-compatible API. The only thing you change between them is the model string in your request.

Sakana Fugu vs Fugu Ultra: How the Two Variants Differ Architecturally

The headline numbers tell part of the story. The architecture tells the rest.

Fugu is tuned for responsiveness. When a request comes in, it tries to resolve as much of the work as possible with a smaller, faster pool of agents. It can still orchestrate (call multiple models, verify, synthesize), but it does so in a more constrained way. For simple requests, it often answers directly without spinning up the full multi-agent loop.

Fugu Ultra is tuned for depth. According to Sakana, Fugu Ultra can coordinate between one and three agents per task, drawing from a deeper expert pool. It is more willing to spend tokens on planning, verification, and synthesis. For hard problems, this means catching errors a single model would miss. For easy problems, it means burning tokens you did not need to burn.

The pool composition is also different. Both variants draw from the same set of frontier models (Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, and others), but Fugu Ultra has access to a deeper roster of specialists for niche tasks. Fugu's pool is optimized for general-purpose performance with predictable latency.

Benchmark Performance Side by Side

Sakana has not published a detailed side-by-side benchmark of Fugu vs Fugu Ultra. The published numbers are Fugu Ultra against external frontier models, including a six-benchmark comparison against Mythos Preview covering SWE-Bench Pro, TerminalBench 2.1, LiveCodeBench, GPQA-Diamond, Humanity's Last Exam, and CharXiv Reasoning. What we know from launch coverage and early hands-on reports:

For most everyday workloads, Fugu produces answers that are functionally equivalent to a single strong model like GPT-5.5 or Opus 4.8, but at a multi-agent quality bar. Fugu Ultra pushes meaningfully beyond what any single model can do, but at the cost of latency and tokens.

Pricing Comparison

Both variants are available through the same billing models, but they behave differently on cost.

Subscription plans (both variants included):

• Standard: $20/month
• Pro: $100/month (10x Standard)
• Max: $200/month (30x Standard)

On subscription plans, you can use either variant. Fugu Ultra will consume your monthly allowance faster because it burns more tokens per query.

Pay-as-you-go API pricing:

The most important nuance: Fugu's pay-as-you-go pricing is dynamic. You pay the standard rate of whichever top-tier underlying model is active for that request. If Fugu routes a query to GPT-5.5, you pay GPT-5.5's rate. If it routes to Opus 4.8, you pay Opus 4.8's rate. Sakana does not stack fees when multiple agents are involved on a single Fugu request.

Fugu Ultra has fixed pricing regardless of which underlying models it activates. This makes Ultra easier to budget for, but it can be more expensive on queries that Fugu would have resolved with cheaper agents.

For the full rate breakdown across both variants and subscription tiers, see our Sakana Fugu pricing guide.

When to Use Fugu

Use Fugu when:

• Latency matters. Interactive chatbots, real-time coding assistants, conversational interfaces where a 30-second response time would feel broken.
• The task is bounded. Single-file code edits, short summaries, focused Q&A, content rewrites. Tasks a single strong model handles cleanly.
• You are running at a high volume. Document classification, content moderation, batch enrichment. Fugu's per-query cost is lower because it does less orchestration work.
• You want intelligent fallback without paying for verification. Fugu still benefits from Sakana's orchestration logic. It can pick the right model for the task. It just does not spend as many tokens checking the answer.

Examples of real workloads that fit Fugu:

• An internal documentation chatbot for your company
• A code review assistant that flags issues but does not have to catch every edge case
• A content generation pipeline producing first-draft marketing copy
• A customer support agent that handles tier-1 questions

When to Use Fugu Ultra

Use Fugu Ultra when:

• The answer matters more than the wait. A wrong answer would have real consequences. A 60-second delay is acceptable.
• The task is genuinely hard. Multi-step coding problems, scientific reasoning, complex analysis where a single model frequently fails or hallucinates.
• Verification is the value. Tasks where catching errors is the whole point: legal document review, security analysis, financial modeling.
• You are working on something high-stakes. Production code that will run unattended, research that will inform business decisions, analysis that will shape strategy.

Examples of real workloads that fit Fugu Ultra:

• A senior engineer's deep-debug session on a gnarly production issue
• Patent prior-art investigation across thousands of documents
• Reproducing a research paper's methodology from scratch
• Kaggle competition entries where every point matters
• Long-horizon agentic tasks running for hours

The Hybrid Pattern Most Teams Should Use

The honest answer for most production deployments is not "Fugu or Fugu Ultra." It is both, routed intelligently.

A pattern that works:

1. Start every request on Fugu. It is faster, cheaper, and good enough for the majority of tasks.
2. Escalate to Fugu Ultra when Fugu's confidence drops or verification fails. If your output includes a self-check or your downstream system can detect bad outputs, use that as a routing signal.
3. Reserve Ultra for explicitly hard tasks. Some workflows you know in advance are hard. Long agent loops, deep research queries, multi-step engineering work. Just send those to Ultra directly.

This blended approach captures most of Ultra's quality lift while keeping your blended latency and cost closer to Fugu's profile.

The Black Box Problem Affects Both

Worth flagging: the routing decisions inside both Fugu and Fugu Ultra are proprietary. You cannot see which specific models handled your query. You cannot reproduce a result by knowing which agents were used.

This applies to both variants equally. The opacity is a feature of Sakana's orchestration architecture, not specific to Ultra. This matters more than it might seem given the export-control disruption to Anthropic's Fable 5 and Mythos Preview that partly motivated Fugu's vendor-independence pitch in the first place. If you are evaluating Fugu for a regulated workload that requires audit trails (healthcare, legal, financial services), this constraint is the same whether you choose Fugu or Fugu Ultra.

If that opacity is a blocker regardless of which variant you pick, these Sakana Fugu alternatives are worth comparing.

Speed vs Quality: A Practical Framework

The decision between Fugu and Fugu Ultra is really a decision about what you are optimizing for in any given workflow.

This framework only works if you have data on your own workload. Sakana's benchmarks are useful directional indicators. Your actual results depend on the specific tasks you are sending and how often the verification rounds in Ultra catch errors that Fugu would have missed.

Building Products on Top of Sakana Fugu

The orchestration capability that Fugu and Fugu Ultra offer is genuinely useful for back-end reasoning and content generation, but most teams shipping AI-powered products today need more than a smart API. They need the full application stack around it: a UI users can actually interact with, a database, authentication, payments, hosting, and the ability to iterate on all of that without a six-person engineering team.

This is where platforms like Emergent close the gap. Emergent is an AI app builder that takes a plain-language description of what you want to build and ships a real, production-ready application end to end. Not a prototype, not a static mockup. A working full-stack product with frontend, backend, database, auth, and deployment handled in a single coordinated pass.

The reason Emergent stands apart from other builders in 2026 is the depth of what it actually generates. Most no-code tools stop at the interface. Emergent reasons through how the whole system should work before building it, then writes real code you fully own. The output syncs directly to your GitHub repository, which means no platform lock-in. You can export it, deploy it elsewhere, or hand it off to an engineering team to extend.

It also handles the messy parts that usually kill AI-built apps. When something breaks, Emergent's multi-agent framework analyzes backend logs and resolves issues without human intervention. When you need to add authentication, payments, or a new integration, you describe what you want and the AI ships it. When your requirements change six months in, you iterate by prompt rather than rebuilding from scratch.

For teams operating in regulated environments, Emergent is SOC 2 Type I certified with SSO/SAML authentication, role-based access control, and audit logging built into the platform. That combination of consumer-grade ease and enterprise-grade compliance is what makes it a genuinely different category from both no-code tools and AI coding assistants.

The point is not that Emergent replaces Fugu or Fugu Ultra. They solve different problems. But if you are using Sakana's orchestration to power smart reasoning inside a product you are building, Emergent is how you ship that product to real users without spending six months on infrastructure first.

The Bottom Line

Fugu and Fugu Ultra are not competing models. They are two settings on the same dial. Fugu is the setting you should default to. Fugu Ultra is the setting you escalate to when the task genuinely warrants it.

For most teams in 2026, the right architecture is hybrid: route the easy stuff through Fugu, escalate the hard stuff to Ultra, and use the savings on volume. If you only need one of the two, pick based on what your workload actually looks like, not on which one has the bigger benchmark numbers.

The framework above is a starting point, not a prescription. The honest way to choose is to run both variants on your actual production tasks for a week, measure latency, cost, and answer quality, and let the data decide.