What is Sakana AI, and how does the Fugu system work?

Sakana AI is a Tokyo-based research lab, and its served lineup is deliberately narrow: it centers on the Fugu multi-agent system rather than a broad tier of standalone LLMs.

Fugu Ultra is its flagship configuration, the one tuned for maximum answer quality by coordinating a deeper pool of expert agents. A few things worth knowing about the lineup:

• Fugu, the system: rather than a single set of weights, Fugu is a learned orchestrator that coordinates a pool of frontier models behind one OpenAI-compatible API, and Fugu Ultra is its highest-quality tier

• Research pedigree: Sakana is best known for research like evolutionary model merging and automated research agents, which is the lineage the Fugu orchestration approach comes out of

• Focused availability: unlike the major labs, Sakana does not offer a wide mini/standard/flagship catalog, so when Fugu Ultra is more than a task needs, you route to models from other providers (see below)

How does Fugu Ultra differ from a conventional single model?

Fugu Ultra is not a single model answering directly. it's an orchestrator that routes a request across a pool of frontier models and composes their work. That shapes where it fits:

• Architecture: it behaves like one model through a single API, but internally coordinates multiple expert agents, which is what drives its quality on hard, multi-step problems

• Quality focus: it's tuned for maximum accuracy and depth on demanding work like AI research, paper reproduction, cybersecurity analysis, and patent or literature investigation, and stands with leading models on rigorous reasoning benchmarks

• Context window: it supports up to 1,000,000 tokens, with pricing that steps up for contexts beyond 272K tokens

• Pricing: at $5 input and $30 output per million tokens, it sits at the premium end, reflecting the multi-agent work happening per request

Fugu Ultra is the right pick when answer quality on a hard, open-ended problem matters more than latency or cost. For routine or high-volume tasks, a single fast model is the better fit.

What models should I consider using alongside Fugu Ultra?

No single model is optimal for every task. Here are models worth pairing with Fugu Ultra depending on what your product needs:

• Gemini 3 Flash: route simple, high-volume, latency-sensitive requests here instead of paying Fugu Ultra's premium for work that doesn't need multi-agent depth

• DeepSeek V4 Pro: cost-efficient frontier reasoning for mid-complexity tasks that sit between "simple" and "needs Fugu Ultra"

• Claude Opus 4.8: an alternative single-model frontier for hard reasoning, useful as a fallback or to sanity-check Fugu Ultra's output on critical work

• GPT-4.1: for agentic, tool-heavy workflows where you want predictable single-model behavior rather than internal orchestration

What are the challenges of using Fugu Ultra in my product?

Like any production LLM, Fugu Ultra comes with tradeoffs worth planning for:

• Premium pricing: at $5 input and $30 output per million tokens, and higher rates above 272K-token context, it's one of the more expensive options, and multi-agent orchestration consumes more tokens per request than a single model

• Latency: coordinating a pool of expert agents adds overhead, so it isn't suited to real-time or interactive use where fast responses matter

• Less predictable behavior: because it decides internally which agents to run, cost and response characteristics per request are harder to forecast than with a single model

• Provider dependency: relying on Sakana alone means an outage or a model change stalls the workloads that depend on it

• Cost at scale: as request volume grows, Fugu Ultra's per-request token use compounds quickly without active budget controls

Why should I use Merge Gateway to route LLM requests with Fugu Ultra and every other model?

Using Fugu Ultra through Merge Gateway gives you access to the model itself and the infrastructure layer around it:

• Cost governance first: given Fugu Ultra's premium pricing, Merge lets you set hard or soft project budgets so spend stays within plan, and attributes every request to a model, project, and tag in one billing dashboard across providers

• Intelligent routing and automatic failover: Merge routes around Sakana outages automatically, and cost, latency, or quality policies can reserve Fugu Ultra for the hard requests while cheaper models take the rest, cutting spend by 40 to 60% with no application code changes

• One API, every provider: reach Fugu Ultra and every other major LLM through a single endpoint and API key, switching providers by swapping the model string

• Build Your Own Router: define what "best" means with curated ML benchmarks or your own eval scores, so the router sends only the requests that justify Fugu Ultra's cost to Fugu Ultra, and explains each decision

• Security and compliance controls: apply DLP rules and prompt injection protection before any request reaches Sakana, and enforce per-project model and region policies outside your application

How can I start routing requests to Fugu Ultra via Merge Gateway?

Getting Fugu Ultra running through Merge Gateway takes a few minutes:

1. Create an account and get your API key from the dashboard.

2. Install the Merge Gateway SDK: run pip install merge-gateway-sdk (Python) or npm install merge-gateway-sdk (Node). Alternatively, if you're already using the OpenAI SDK, set base_url = "https://api-gateway.merge.dev/v1/openai" and your existing code works as-is.

3. Make your first request using the provider/model format. For Fugu Ultra, the model string is sakana/fugu-ultra. Swap the model string to route to any other provider without changing anything else.

4. Configure a routing policy in the dashboard to set failover behavior, cost limits, and optimization strategy. Your first policy can be as simple as naming Fugu Ultra as primary with one fallback.

Full setup instructions and SDK references are in the Merge Gateway docs.