Complex coding and review
Use a coordinated pool for difficult implementation, debugging, test, and code-review tasks where a single pass is often insufficient.
Independent tool overview
Sakana Fugu at a glance
Sakana Fugu is a hosted multi-agent orchestration service that looks like one language model API. Behind the endpoint, a learned coordinator decides whether to answer directly or route work among multiple foundation models, assigning, checking, and combining their contributions for complex coding, research, reasoning, and security tasks.
Visit the official Sakana Fugu site ↗
- Product type
- Hosted multi-model orchestration API
- Models
- Fugu, Fugu Ultra, and Fugu Cyber
- API compatibility
- OpenAI Chat Completions and Responses
- Availability
- Generally available, excluding EU/EEA
- Starting subscription
- $20 per month
- Last reviewed
- August 29, 2026
Overview
What Sakana Fugu is
Fugu aims to remove the application-level complexity of building a multi-model agent system. Developers call an OpenAI-compatible Chat Completions or Responses endpoint, while Sakana's coordinator handles model selection, delegation, verification, and synthesis. The system builds on Sakana AI's ICLR 2026 TRINITY and Conductor research.
The current line has three variants. Fugu balances latency and quality for everyday coding, review, chat, and analysis. Fugu Ultra uses a deeper, fixed agent pool to maximize answer quality on difficult multi-step work. Fugu Cyber is specialized for security reasoning and is sold through a contact-sales process.
Multi-model routing can diversify the workers behind an answer, but it does not eliminate vendor dependence: customers still rely on Sakana's proprietary coordinator, endpoint, policy, and model-pool decisions. Fugu does not expose which models were used for a request, and Ultra does not allow provider opt-outs, so regulated teams should evaluate routing opacity, data residency, auditability, and geographic availability before adoption.
Use cases
Who Sakana Fugu is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Long-form technical research
Delegate multi-step literature, patent, scientific, and paper-reproduction work that benefits from planning, specialist calls, and verification.
Model-diverse applications
Give an application access to multiple underlying model strengths through one familiar API without hand-building a router and agent graph.
Specialized security analysis
Evaluate Fugu Cyber for authorized vulnerability research and threat-analysis work where a security-specific coordinator is useful.
Capabilities
Core Sakana Fugu features
Learned orchestration
Uses trained coordinator models rather than only static routing rules to decide when to delegate and how workers should communicate.
One OpenAI-compatible API
Exposes Chat Completions and Responses-style endpoints so many existing clients and coding harnesses can switch base URL and credentials without a new SDK.
Dynamic model pool
Fugu can select, assign, verify, and synthesize work from different foundation models, including recursive calls to the Fugu system itself.
Fugu
The balanced default emphasizes useful quality with lower latency for interactive coding, review, analysis, and chatbot workloads.
Fugu Ultra
Coordinates a deeper, fixed expert pool for difficult, high-stakes, and long-running work where answer quality matters more than response time.
Fugu Cyber
Specializes the orchestration approach for security analysis, vulnerability research, and threat-investigation tasks.
Provider opt-outs in Fugu
The standard Fugu model lets customers exclude specific providers or models to address some privacy, compliance, or organizational requirements.
Codex and Claude Code launchers
Sakana's public setup project configures Fugu as the model backend for supported Codex and Claude Code workflows on macOS and Ubuntu.
Process
How the Sakana Fugu workflow works
Step 1
Classify the workload
Use Fugu for interactive or cost-sensitive work, Ultra for difficult multi-step problems, and Cyber only for authorized security use cases after commercial review.
Step 2
Define data constraints
Decide which information can be sent, whether training-data use must be disabled, and whether any provider must be excluded before creating production credentials.
Step 3
Integrate the compatible endpoint
Point a supported OpenAI client or coding harness to Sakana's endpoint and select a current Fugu model identifier.
Step 4
Run a representative evaluation
Compare accuracy, latency, token consumption, cost, and failure modes against the strongest single model on your own tasks rather than relying on vendor benchmarks.
Step 5
Constrain agent actions
Put approvals around file writes, deployment, security testing, external communication, spending, and any other action with material consequences.
Step 6
Monitor request-level cost
Capture the token and price report for each request, set budgets and timeouts, and route routine work away from Ultra when the added depth is not valuable.
Cost
Sakana Fugu pricing and free plan
Sakana offers monthly subscriptions and pay-as-you-go tokens. Standard is $20 per month, Pro is $100 with 10 times the Standard allowance, and Max is $200 with 20 times the allowance; all three include Fugu and Fugu Ultra, but Sakana does not publish the numeric Standard token allowance on the product page. Pay-as-you-go Fugu charges one rate based on the highest-tier active underlying model rather than summing every worker call. Fugu Ultra v1.1 and v1.0 cost $5 per million input tokens, $30 per million output tokens, and $0.50 per million cached-input tokens, increasing to $10, $45, and $1 above 272K context. Fugu Cyber pricing is quote-based.
Standard subscription
$20/month
For lightweight daily use, occasional API calls, and small personal experiments.
- Includes Fugu and Fugu Ultra
- Baseline usage allowance
- Exact numeric allowance is not published on the product page
Pro subscription
$100/month
For regular coding, review, research, and analysis sessions.
- Includes Fugu and Fugu Ultra
- 10x the Standard usage allowance
- Lower priority than pay-as-you-go tokens
Max subscription
$200/month
For power users running deeper or longer sessions.
- Includes Fugu and Fugu Ultra
- 20x the Standard usage allowance
- Lower priority than pay-as-you-go tokens
Fugu pay as you go
Underlying-model rate
Charges a single blended rate based on the highest-tier model in the active pool.
- Model fees are not stacked
- Higher service priority than subscription tokens
- Request-level usage and cost are reported
Fugu Ultra pay as you go
$5 input / $30 output per 1M tokens
Fixed token rates for current Fugu Ultra v1.1 and v1.0.
- $0.50 per 1M cached-input tokens
- Above 272K context: $10 input, $45 output, $1 cached input per 1M tokens
- Long orchestration runs can still generate substantial output cost
Fugu Cyber
Contact sales
Specialized security model with commercial terms provided by Sakana.
- No public rate card
- Confirm authorized-use, support, privacy, and deployment terms with sales
Pricing checked . Check current pricing at the source ↗
Assessment
Sakana Fugu strengths and limitations
Where it stands out
- Hides multi-model planning, routing, verification, and synthesis behind a familiar API
- Standard, Ultra, and Cyber variants make the intended workload tradeoff clearer
- Fugu can exclude selected workers for some data and compliance requirements
- Pay-as-you-go orchestration does not simply stack every underlying model's token bill
- Subscription and token pricing are publicly described in useful detail
- The architecture is grounded in published orchestration research
What to consider
- Sakana does not reveal which underlying models handled an individual request or how they were coordinated
- Fugu Ultra uses a fixed full pool and does not support the provider opt-outs available in standard Fugu
- The product is not offered to users in the EU or EEA while Sakana works toward regulatory compliance
- The numeric monthly token allowance behind the Standard, Pro, and Max multipliers is not published on the marketing page
- Vendor benchmark and testimonial results should be reproduced on the buyer's own workloads before supporting a production claim
- Multi-model orchestration can increase latency, output volume, complexity, and cost on tasks that a single strong model could solve
- Usage data may be used to improve Fugu unless the customer opts out in the console
Compare
Sakana Fugu alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Agents
Perplexity Computer
Consider Perplexity Computer when you want a finished multi-model agent product for long-running tasks rather than an API-first orchestration model.
Explore Perplexity Computer →Consumer
GPT 5.5
Consider GPT-5.5 when one frontier model and its native agent ecosystem are simpler and more transparent for the workload.
Explore GPT 5.5 →Consumer
Gemini 3.1 Pro
Consider Gemini 3.1 Pro when Google's multimodal context, tooling, and platform integrations matter more than hidden multi-model routing.
Explore Gemini 3.1 Pro →Consumer
Claude Opus 4.8
Consider Claude Opus 4.8 for direct access to a high-capability model with strong coding and agentic workflows.
Explore Claude Opus 4.8 →Questions
Sakana Fugu FAQs
What is Sakana Fugu?
Sakana Fugu is a hosted multi-agent system exposed as one OpenAI-compatible language-model API. A learned coordinator selects and manages multiple underlying models, then returns one synthesized response.
What is the difference between Fugu and Fugu Ultra?
Fugu balances quality and latency for everyday work and allows selected provider opt-outs. Fugu Ultra uses a deeper fixed pool to maximize quality on complex, multi-step tasks, with greater latency and no worker opt-outs.
How much does Sakana Fugu cost?
Monthly subscriptions are $20 Standard, $100 Pro, and $200 Max. Pay-as-you-go Fugu uses a single rate based on the highest-tier active worker. Fugu Ultra is $5 per million input tokens, $30 output, and $0.50 cached input, with higher rates above 272K context. Cyber is quote-based.
Does Fugu reveal which models it used?
No. Sakana says the selected workers and coordination strategy are proprietary and are not exposed per query.
Can I exclude a model provider?
Standard Fugu lets users opt specific models or providers out in the console. Fugu Ultra requires its fixed full pool and does not provide the same control.
Does Sakana train on customer usage?
Sakana says usage data can help improve Fugu, but users can opt out of training-data use at any time through the console. Confirm contract terms for sensitive enterprise data.
Is Sakana Fugu available in Europe?
No. As of August 29, 2026, Sakana says Fugu is not available to users in EU or EEA member states while it works toward GDPR and regional compliance.
Bottom line
Our Sakana Fugu verdict
Sakana Fugu is an interesting choice for teams that want the potential benefit of multiple frontier models without maintaining their own coordinator, router, verifier, and provider integrations. The honest tradeoff is opacity: buyers cannot audit per-request routing, Ultra's pool is fixed, and Sakana becomes the new control-plane dependency. Run a costed evaluation on your own hard tasks, enable the data-use opt-out when appropriate, and use Standard Fugu's provider controls before moving regulated or production work into the system.
Visit Sakana Fugu website ↗Similar Consumer Tools
View all Consumer AI tools →
Framer 3.0
Framer 3.0 - Canvas-native agents that build and run your whole site
Claude Sonnet 5
Claude Sonnet 5 - Anthropic's mid-sized, cost-effective model with strong agentic and coding capabilities
Fusion
Fusion - OpenRouter's new tool that fuses several models into panels to achieve frontier performance
Higgsfield Explainer
Higgsfield Explainer - Turn any topic into a faceless explainer video
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.