Sakana Fugu: The Model Router Hiding Inside One API

Sakana Fugu is a multi-agent orchestration system wrapped as one model API.

That is the important part.

You do not manually call five different models.
You do not build your own routing layer from scratch.
You do not hard-code “use Model A for coding, Model B for reasoning, Model C for long context.”

You send a request to Fugu.

Fugu coordinates the model pool behind the scenes.

Verified: Sakana describes Fugu as a system that dynamically orchestrates “the world’s best models” through a single API.

The clean way to think about it:

That is why this launch matters.

The model market is crowded. The routing layer is still early.

The beginner question is:

The better builder question is:

That is the real shift.

Modern AI work is no longer one prompt, one answer.

Real workflows look like this:

One model can do all of that badly.

A multi-agent setup can split the job.

Sakana’s pitch is that Fugu learns how to assemble and coordinate those agents instead of relying on hand-designed workflows.

That is the unlock.

Not more prompts.

Better delegation.

Fugu can replace part of your manual model-selection workflow.

You decide:

You call one API.

Fugu handles model selection and switching.

Verified: Sakana says Fugu provides access to a coordinated pool of specialized models through one API and handles model selection and switching for each task.

What it does not replace:

Wrong assumption: orchestration means you can stop thinking.

No.

It means the model stack gets more automated, but your workflow still needs guardrails.

Sakana offers two models:

Verified: Sakana says Fugu balances latency and quality, while Fugu Ultra prioritizes answer quality on complex multi-step work.

Use Fugu when speed matters.

Use Fugu Ultra when getting the answer right matters more than waiting.

That is the practical split.

Sakana makes aggressive benchmark claims.

Their table shows Fugu and Fugu Ultra compared against Opus 4.8, Gemini 3.1 Pro, and GPT 5.5 across tasks like:

Source claim: Sakana says Fugu models surpass publicly accessible frontier models and sit close to Fable 5 and Mythos Preview across engineering, scientific, and reasoning benchmarks.

Important:

Sakana says Fable 5 and Mythos Preview are not in Fugu’s agent pool because they are not publicly accessible.

That matters because the claim is not:

The claim is:

That is a very different statement.

Source claim: These numbers are from Sakana’s own benchmark table, not independent testing.

Use the numbers carefully.

Strong hook.

Not final proof.

Sakana offers subscription and pay-as-you-go pricing.

Verified: Every subscription tier includes both Fugu and Fugu Ultra.

For fugu-ultra-20260615:

For context above 272K tokens:

Verified: These prices are listed on Sakana’s Fugu pricing page.

Sakana says Fugu does not stack model fees when multiple agents are active.

Instead, you pay one rate based on the top-tier model involved.

That is important.

Multi-agent systems can get expensive fast when every agent call adds another provider bill.

Fugu’s pitch is that the pricing does not multiply just because more agents are active.

Still, watch your token usage.

More orchestration can mean more tokens.

More tokens still means more cost.

Sakana says Fugu is available through an OpenAI-compatible API.

That means you should be able to point an existing client or coding harness at the Fugu endpoint using an API key, without an SDK migration.

Sakana’s public page confirms OpenAI-compatible access, but the exact command template was not visible in the accessible page.

Do not invent it.

No.

This is closer to a model manager.

Fugu coordinates multiple agents. The orchestration is the product.

No.

Benchmarks are controlled tests. Production workloads involve messy inputs, latency limits, budget limits, compliance rules, and failure recovery.

No.

Fugu Ultra is for harder work where quality matters more than speed.

For normal interactive work, Fugu is probably the cleaner first test.

No.

Sakana says the specific models selected and coordination method are proprietary and not exposed.

That can be a dealbreaker for teams that need routing transparency.

No.

Sakana says Fugu is not available in the EU/EEA right now while it works toward GDPR and EU-specific compliance.

For a Berlin-based user, this matters immediately.

You cannot see the exact underlying model route.

That makes debugging harder.

If the answer fails, you may not know whether the issue came from planning, execution, verification, or a specific model in the pool.

Fugu lets users opt out of specific models for the regular Fugu model, but Fugu Ultra uses the full fixed agent pool.

If your company has provider restrictions, check this before building around Ultra.

More agents usually means more coordination.

More coordination usually means slower responses.

Sakana directly says Fugu Ultra prioritizes answer quality at the cost of response time.

The pricing looks clean, but long-running agentic jobs can burn tokens quickly.

Especially with:

EU/EEA access is blocked right now.

That alone makes it unusable for a chunk of builders until compliance changes.

Use Fugu like an evaluation layer, not blind infrastructure.

Good first tests:

Avoid first tests like:

Run the same task through:

Score them on:

Do not waste Ultra on simple tasks.

Use it for:

Fugu is still AI output.

Treat it like a powerful analyst, not a final authority.

By The AI Leverage - Learn and master AI daily