How to Use Sakana AI Fugu: From API Key to Claude Code | Pricing, Free Tier, and Performance

Image source: Sakana AI, "Sakana Fugu" official page. All figures in this article are quoted from Sakana AI's official site.

On June 22, 2026, Sakana AI made Fugu generally available — an AI platform that dynamically orchestrates multiple large language models (LLMs) and exposes them through a single API. Rather than relying on one frontier model, it conducts several of the world's leading models internally and behaves as though it were a single model.

This article walks through the concrete steps from issuing an API key to your first request, how to drive Codex and Claude Code with Fugu, how to choose between Fugu, Fugu Ultra, and Fugu Cyber, what it costs and whether a free tier exists, and how it performs on benchmarks — all based on the official documentation.

---

What is Sakana Fugu?

Sakana Fugu is a multi-agent orchestration platform built by Sakana AI. To the user it is a single API — an OpenAI-compatible one at that — but internally it calls specialized models (agents) dynamically according to the task, handling delegation, verification, and integration automatically.

• Used like a single model: model selection and delegation are handled internally, so none of that complexity reaches the caller.
• Dynamically bundles models: it selects the best model from a pool of agents, and may recursively invoke Fugu itself when needed.
• Avoids vendor lock-in: if one provider restricts access, it reroutes to another model and continues.
• Stable persona across long sessions: it maintains consistent responses and persona through extended exchanges.

Fugu reached general availability on June 22, 2026, following a beta with roughly 500 early users. Today the lineup centers on Fugu, with purpose-built models available under the same API key.

---

How to use Fugu: from API key to first request

Everything starts at console.sakana.ai. Issue an API key there, then simply repoint your OpenAI-compatible client.

Source: Sakana AI Console, "Get started"

1. Create an API key

Log in to the console and create an API key. The generated key is shown only once, so copy and store it before closing the dialog. You also choose the billing mode (subscription or pay-as-you-go) per key at this point.

Enabling Fugu custom model pool when creating the key lets you narrow which providers Fugu routes to. If data handling or compliance requirements rule out certain providers, exclude them here. Leave it off to use the default pool.

2. Verify connectivity with curl

All API endpoints live at https://api.sakana.ai. Start with a connectivity check.

export SAKANA_API_KEY={your api key}

curl -X POST https://api.sakana.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $SAKANA_API_KEY" \
  -d '{"model":"fugu","messages":[{"role":"user","content":"How many r in word strawberry"}]}'

3. Call it from the OpenAI SDK

Fugu supports the standard OpenAI SDK interface, so swapping base_url is enough to reuse existing code. For generation requests, the vendor strongly recommends the Responses API for better performance.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.sakana.ai/v1",
    api_key="YOUR_API_KEY",
)

response = client.responses.create(
    model="fugu-ultra",
    input="Write a concise explanation of how TLS works",
    timeout=120.0,
)

print(response.output_text)

Supported endpoints are the OpenAI-compatible Chat Completions API, Responses API, and Models API, plus the Anthropic-compatible Messages API.

4. Control reasoning effort

Fugu exposes a reasoning.effort parameter to control how deeply it thinks. The accepted values differ by model.

max functions as a distinct maximum level only on fugu-ultra-v1.1, reserved for the hardest problems. Any other value is rejected.

Note that temperature, top_p, stop, and seed are accepted but ignored. previous_response_id is not accepted at all, so the full conversation history must be sent in input on every turn.

5. Use built-in tools such as web search

The Responses API supports OpenAI-compatible built-in tools. Just add web_search to the tools array.

response = client.responses.create(
    model="fugu",
    tools=[{"type": "web_search"}],
    input="Search the web for today's top AI news and summarize it with citations.",
)

---

Driving Codex and Claude Code with Fugu

The most practically significant capability is that you can swap the engine behind a coding agent for Fugu. Official instructions exist for two: Codex CLI and Claude Code.

One-line installation

The simplest route is the official installer (Ubuntu and macOS).

curl -fsSL https://sakana.ai/fugu/install | bash

After installing, start each tool through its dedicated launcher.

codex-fugu   # start Codex on Fugu
claude-fugu  # start Claude Code on Fugu

On Windows, or where the installer does not complete, set it up manually instead.

Configuring Claude Code manually

You can point Claude Code at Fugu with environment variables alone, without the launcher. Note that you use ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY.

export ANTHROPIC_BASE_URL="https://api.sakana.ai"
export ANTHROPIC_AUTH_TOKEN="fish_..."            # your Sakana API key
export ANTHROPIC_DEFAULT_OPUS_MODEL="fugu-ultra[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL="fugu[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="fugu[1m]"
export CLAUDE_CODE_SUBAGENT_MODEL="fugu[1m]"

claude

The Opus tier maps to fugu-ultra[1m], and Sonnet and Haiku both map to fugu[1m]. The settings apply only to the current shell session.

Can you use it from Cursor or other tools?

Official instructions cover only Codex and Claude Code, but because Fugu is served as an OpenAI-compatible API, any tool that lets you set an arbitrary base URL and API key can connect the same way — Cursor's custom OpenAI-compatible model registration being the obvious case. There is no official support statement or setup guide for it, so treat it accordingly.

---

The model lineup and when to use each

The same API key gives you access to all of the following.

Fugu suits embedding into tools like Codex for coding and code review, or running a responsive chatbot. Fugu Ultra coordinates a broader pool of specialist agents to maximize answer quality on hard, high-stakes problems — Kaggle competitions, paper reproduction, cybersecurity analysis, literature and patent research.

Fugu Cyber targets workflows where depth and precision matter: security analysis, vulnerability research, threat detection. Access requires submitting an application form stating your intended use and verifiable contact information, which Sakana AI reviews and approves manually. The vendor reports success rates of 86.9% on CyberGym and 72.1% on CTI-REALM, calling it comparable to leading security-specialized frontier models such as GPT-5.5-Cyber and Mythos-Preview.

Sakana Namazu is a separate line: an LLM specialized for Japanese, with web search and code execution as built-in tools. It also supports image input, file input (PDF, XLSX, CSV, DOCX and more), and extended thinking. It is excluded from the subscription plans and available on pay-as-you-go only.

---

Fugu pricing: is there a free tier?

The short answer first: Fugu has no free tier — no free trial and no starter credits. Using it requires either a subscription plan or buying tokens on pay-as-you-go.

Subscription plans

Source: Sakana AI Console, "Pricing"

Every plan includes both Fugu and Fugu Ultra. The plans do not differ in which models you can reach — only in the size of the allowance.

Pay-as-you-go (Fugu Ultra)

Fixed rates for fugu-ultra-v1.1 and fugu-ultra-v1.0, per million tokens.

fugu-cyber-v1.0 runs about 20% higher: $6 input, $36 output, $0.60 cached input (each doubling above 272K). sakana-namazu-v1.0 is far cheaper at $0.95 input, $4.00 output, and $0.15 cached input, with web search billed separately at $7.00 per 1,000 calls and code execution at $0.12 per hour.

Also worth knowing: tokens purchased on pay-as-you-go are processed at higher priority than monthly-plan tokens. Pay-as-you-go is recommended for production workloads where reliability matters.

---

Fugu benchmark performance

The table below is the official benchmark comparison of Fugu / Fugu Ultra against Opus 4.8, Gemini 3.1 Pro, and GPT 5.5. Fugu's scores are from the v1.0 models.

It posts top scores on many items, including GPQA-D (95.5) and LiveCodeBench (93.2), and on SWE Bench Pro Fugu Ultra's 73.7 exceeds Opus 4.8's 69.2. What stands out is that an approach built on bundling models matches or beats single frontier models.

Table source: Sakana AI Console, "Models"

---

How Fugu works (multi-agent orchestration)

As the opening diagram shows, Fugu conducts a pool of LLMs — closed and open models, plus Fugu itself — and routes each task to the best-suited model. At its core are two papers Sakana AI presented at ICLR 2026.

• Trinity: a lightweight, evolutionarily optimized "coordinator" directs multiple LLMs across several turns, assigning roles such as Thinker, Worker, and Verifier and delegating adaptively by task.
• Conductor: trained with reinforcement learning, it discovers coordination strategies in natural language on its own, learning non-obvious but efficient patterns rather than relying on hand-designed workflows.

Fugu itself is a language model specialized in understanding when to delegate and how to combine specialist outputs.

Related research: Sakana Fugu technical report (arXiv:2606.21228), Trinity (arXiv:2512.04695, ICLR 2026), Conductor (arXiv:2512.04388, ICLR 2026)

---

Why Fugu matters right now

Fugu's timing reflects the fact that the supply of frontier models has begun to face geopolitical risk. Anthropic's latest models, Fable 5 and Mythos 5, were abruptly cut off in June 2026 by US export-control direction (see the Claude Fable 5 article for details).

Fugu positions itself as a practical answer to that single-vendor dependency risk. Because the design bundles multiple models, it can reconfigure the agent pool and hold frontier-class performance even when one provider becomes unavailable — the idea Sakana frames as "AI sovereignty."

---

Real use cases and user feedback

During the beta, Fugu was applied to automated data science research, paper reproduction, cybersecurity analysis, code review, and patent and literature searches. A few officially published comments:

• Code review: "In code review, Fugu Ultra was clearly better than GPT-5.5. Where a competitor found only three issues, Fugu flagged more than twenty." (software engineer)
• Stability over long sessions: "Output quality matched the top frontier models. On top of that, the persona stayed remarkably stable across long sessions." (head of an enterprise platform)
• Security assessment: "From a single scope instruction, Fugu ran a security assessment end to end — reconnaissance, XSS/SQLi checks, authentication review, and report writing." (security engineer)

Other reports have it outperforming Gemini 3.1 Pro, Opus 4.8, and GPT 5.5 on multi-step hard tasks including generating a Rubik's Cube solver, mechanical CAD design, blindfold chess at expert level, and stock trading analysis (achieving an average return of +19.43%).

You can also see an overview of Sakana Fugu on the card below.

---

Things to be aware of

• Regional availability: while work continues toward GDPR and EU-specific compliance, service is not provided to users in EU or EEA member states.
• Model update cadence: when a new frontier model is released publicly, Sakana expects to spend roughly two weeks training and evaluating updated Fugu models before rolling them out.
• Fugu Cyber is gated: you must submit an application form with your intended use and verifiable contact details and pass a manual review.
• Send the whole history every turn: previous_response_id is not supported, so multi-turn conversations require including the full history in input.

---

Frequently asked questions about Sakana Fugu

Does Sakana Fugu have a free tier?

No. Using it requires either a subscription plan (Standard $20 / Pro $100 / Max $200 per month) or buying tokens on pay-as-you-go. If you only want to try it, the $20 Standard plan is the cheapest entry point.

Is Sakana Fugu a single model?

From the user's perspective it behaves as one API and one model, but internally it is a multi-agent platform that dynamically bundles and coordinates multiple specialist LLMs.

How do Fugu and Fugu Ultra differ?

Fugu balances performance and latency for everyday work; Fugu Ultra coordinates a broader pool of specialist agents to maximize answer quality on hard problems. Both are available on every subscription plan. Reasoning depth differs too: only fugu-ultra-v1.1 accepts max as a distinct maximum level.

Which model ID should I specify?

Use fugu for everyday work and fugu-ultra (v1.1 by default) for hard problems. To pin the older version, specify fugu-ultra-v1.0 (also known as fugu-ultra-20260615). There is also the security-focused fugu-cyber and the Japanese-specialized sakana-namazu.

Can I use it from Cursor or Claude Code?

Claude Code is officially supported: install with curl -fsSL https://sakana.ai/fugu/install | bash and launch with claude-fugu. Codex works the same way via codex-fugu. There is no official procedure for Cursor, but since Fugu is an OpenAI-compatible API, any tool that lets you set the base URL (https://api.sakana.ai/v1) and an API key can connect on the same principle.

Which APIs is Fugu compatible with?

The OpenAI-compatible Chat Completions API, Responses API, and Models API, plus the Anthropic-compatible Messages API. For generation requests the vendor strongly recommends the Responses API on performance grounds.

Does running multiple agents cost more?

Model charges are not stacked; you pay a single rate based on the highest-tier model involved. That said, on Fugu Ultra the tokens spent on orchestration are included in usage and count toward the final price, at the same rate as ordinary input and output tokens.

Is it available in Japan?

Yes. The only exclusion is users in EU and EEA member states; there are no restrictions on use from Japan.

---

Summary

• Sakana Fugu is a multi-agent platform that dynamically bundles multiple LLMs behind a single OpenAI-compatible API (generally available since June 22, 2026)
• To use it, create an API key at console.sakana.ai and repoint base_url to https://api.sakana.ai/v1. Codex and Claude Code are covered by a one-line installer
• The models are fugu, fugu-ultra, fugu-cyber (by application), and the Japanese-specialized sakana-namazu — all under the same API key
• There is no free tier; choose a subscription ($20 / $100 / $200) or pay-as-you-go (Fugu Ultra at $5 input and $30 output per million tokens). Max's allowance is 20x Standard
• Built on the Trinity and Conductor research (ICLR 2026), it matches or beats single frontier models with SWE Bench Pro 73.7, GPQA-D 95.5, and LiveCodeBench 93.2

Fugu's approach of bundling models is a genuinely new idea — worth putting to work in your own products and workflows.