Sakana Fugu
Sakana AI · Coding
Sakana Fugu is a multi-agent AI model orchestration system from Sakana AI that ships as a single OpenAI-compatible API. Instead of training one bigger model, it learns to assemble and coordinate a pool of specialized agents and hands you the combined result as if you called a normal model. Think of it as a model orchestration API you point at once. You pick a tier (Fugu, Fugu Ultra, Fugu Max, or Fugu Cyber), send your usual request, and let the system decide which models do the work. The company bills it as a frontline AI API for complex, multi-step jobs.
About Sakana Fugu
What Is Sakana Fugu
Sakana Fugu takes a different bet than most model launches. Rather than building the largest single model, Sakana AI trained a coordinator that learns how to route, delegate, and verify work across many strong models. From the outside it looks like one model; on the inside, a pool of agents splits a task, runs the pieces, checks the results, and merges everything into one answer.
The pitch is vendor independence. Because the pool is swappable, a single provider shutting an API off doesn't take the whole system down. That matters for teams that build agent products and can't afford a sudden cutoff. The trade-off is transparency: Fugu doesn't tell you which underlying models handled a given query, since that routing logic is the company's core research.
So what does that mean in practice? If a result looks wrong, you can't tell which agent produced it. Debugging gets harder. No clean fix.
It also won't work everywhere. Sakana AI doesn't serve users in the EU or EEA while it works on GDPR compliance, so businesses in those regions need another option for now.
Getting Started
- Get an API key from the Sakana AI console and pick your tier (Fugu for balance, Fugu Ultra for quality, Fugu Max for cost-performance, Fugu Cyber for security work).
- Point your existing OpenAI client or coding tool at the Fugu endpoint. No SDK migration is needed.
- Send a request the way you normally would, using the model name you chose.
- If you're on Fugu, open the console settings to exclude specific providers or models for privacy or compliance reasons.
- Watch token usage and cost per request in the console, and scale up once you've seen the numbers.
Product Information
A quick look at Sakana Fugu's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- Developers building agent products
- Software teams doing heavy code review
- Researchers and analysts
Tasks
- Multi-step coding and reasoning
- Autonomous research
- Cybersecurity analysis
- Batch production workloads
Scenarios
- You want to stop depending on one model vendor
- A team needs governance over which models see its data
- You're prototyping an agent feature and don't want to build a router
Key features
One API, Many Models
Fugu exposes a coordinated pool of specialist models behind a single, OpenAI-compatible endpoint. Your integration doesn't change when the underlying pool does, and you can switch between Fugu, Fugu Ultra, and Fugu Max by changing one model name. Sakana AI positions this as reducing API complexity while improving cost-performance. Simple idea. Real payoff.
Learned Orchestration, Not Hand-Written Workflows
Most agent frameworks make you hard-code which model does what. Fugu learns those patterns instead. Its coordination is grounded in two ICLR 2026 papers, TRINITY and the Conductor, which show a system assembling and routing expert agents per task rather than following a fixed flowchart. Not a fixed pipeline. A learned one.
Four Tiers for Different Jobs
Fugu balances performance and latency for everyday work like coding and chatbots. Fugu Ultra coordinates a deeper pool for hard, high-stakes problems such as Kaggle competitions and paper reproduction. Fugu Max orchestrates the largest pool to push cost-performance, and Fugu Cyber is tuned for security reasoning. According to Sakana AI, Fugu Cyber scores 86.9% on CyberGym and 72.1% on a second industry security benchmark.
Agent Opt-Out and Vendor Independence
On the standard Fugu tier, you can exclude specific models or providers from the pool through console settings. This is the feature that separates Fugu from a plain router: it lets teams meet data, privacy, and compliance requirements without leaving the product. Enterprise customers can request custom model configurations directly.
Transparent Per-Request Cost
Token usage and cost come back with each request, so you can monitor spend in real time and forecast before scaling up. When several agents are active in a request, you're billed a single blended rate based on the top-tier model involved instead of stacking every model's fee.
No SDK Migration
Because the API is OpenAI-compatible, you point your current client at the Fugu endpoint and start sending requests. There's no new SDK to learn and no rewrite of your existing chat or coding tool. That's the whole setup.
Pros and cons
Pros
- One OpenAI-compatible endpoint hides multi-model routing, so integration stays simple.
- Swappable model pool reduces the risk that one vendor's API shutdown breaks your product.
- Agent opt-out on the standard tier supports privacy and compliance needs.
- Published benchmark numbers for the specialized Fugu Cyber tier give security teams something concrete to compare.
- No stacking of model fees keeps multi-agent billing easier to predict.
Cons
- You can't see which models handled a query, so debugging unexpected output is harder. If a result is wrong, you don't know which agent to blame.
- No service in the EU or EEA yet, which rules it out for users in those regions until compliance work finishes.
- Fugu Ultra and Fugu Max have fixed pools you can't trim, so teams needing strict model control have to fall back to the standard tier.
- Long-running autonomous jobs consume tokens quickly, and the higher Ultra rates ($30 per 1M output tokens) add up fast.
Frequently asked questions
It's a multi-agent orchestration API for complex, multi-step tasks such as coding, reasoning, research, and security analysis. You call it like a normal model, but it coordinates several specialist models to produce the answer.
Related content
Explore related tools, skills, and articles for Sakana Fugu.
Sakana Fugu Alternatives
Base Code
Base44 (Wix) · Coding · ProductivityBase Code is a shared cloud development environment from Base44. Teammates describe changes to an existing GitHub codebase in chat, preview them live and send them back as pull requests.
Google EmbeddingGemma 2
Google DeepMind · Writing · Image · CodingGoogle EmbeddingGemma 2 is a 740M-parameter open model that turns text, code, images, audio, and video frames into vectors in one 768-dimension space, running on phones and laptops.
GitGlow
GitGlow · Coding · ProductivityGitGlow is a local-first desktop Git client that lets independent developers review their own code and their AI agents' changes in one place before shipping.