Fugu on OpenRouter: Slugs, Pricing Parity, and the Catches

OpenRouter is the fastest way to call Sakana Fugu without a Sakana account: same models, same per-token prices, one unified billing account, and keys with hard credit limits — the closest thing to a spending cap that exists for Fugu Ultra. It’s also the route with real behavioral differences from the direct API, some of which will break assumptions. This guide covers the exact model slugs, verified pricing, working code, and every documented difference. Verified against OpenRouter’s live model listings on August 18, 2026, with prices re-checked against Sakana’s on August 19.

What OpenRouter is, in one paragraph

OpenRouter is an API aggregator: one account, one key format, one OpenAI-compatible endpoint, hundreds of models from many providers behind it. You pay OpenRouter, OpenRouter pays the provider. For Fugu specifically, requests are served by a single first-party endpoint — Sakana AI itself — so you’re getting the real service, brokered. Sakana lists OpenRouter among its official distribution partners.

Which Sakana models are on OpenRouter

Exactly two, as of August 18, 2026 — and the slugs are easy to get wrong:

Model	Slug	Permanent slug
Fugu Ultra	sakana/fugu-ultra	sakana/fugu-ultra-20260615
Sakana Namazu	sakana/sakana-namazu (note the doubled “sakana”)	sakana/namazu-20260811

Base Fugu is not listed. If your workload wants the cheaper, latency-balanced fugu model, that requires the direct Sakana API — on OpenRouter, the Fugu family means Fugu Ultra. (Namazu, remember, is a separate Japanese-specialized model, not a Fugu variant — but it’s on OpenRouter at a tenth of the cost and makes a good pipeline-testing stand-in.)

Pricing: identical to the direct API

We compared OpenRouter’s listed rates against Sakana’s official pricing line by line. They match exactly, including the tiers (as of August 2026): Fugu Ultra at $5 input / $30 output / $0.50 cached input per million tokens, switching to $10 / $45 / $1.00 above 272K prompt tokens; Namazu at $0.95 / $4.00 / $0.15. OpenRouter’s model metadata even encodes the 272K threshold as an explicit pricing override — so the long-context tier follows you here.

So the routing decision isn’t about price. It’s about account logistics (no Sakana account needed, unified billing across many models, per-key credit limits) versus feature completeness (base Fugu, the custom model pool, and subscription plans exist only on the direct API).

Working code

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="sk-or-...",   # your OpenRouter key
)

response = client.chat.completions.create(
    model="sakana/fugu-ultra",
    messages=[{"role": "user", "content": "Explain how TLS works, concisely"}],
    reasoning_effort="high",   # high | xhigh | max
    timeout=120.0,
)
print(response.choices[0].message.content)

Same OpenAI SDK, different base URL and key format. Keep the generous timeout habit from the direct-API guide — the model behind the broker is the same multi-agent system with the same long turns.

The differences that will actually affect you

All of the following comes from OpenRouter’s own model metadata for sakana/fugu-ultra, checked August 18, 2026:

• Reasoning is mandatory and defaults to xhigh. You cannot disable it, and the supported efforts are high, xhigh, and max. Note that default: left alone, OpenRouter runs Fugu Ultra at xhigh, one notch above the direct-API default — set high explicitly if you want the cheaper baseline. (Our playground defaults to high for exactly this reason.)
• Sampling parameters aren’t supported — temperature, top_p, and max_tokens aren’t in the supported-parameter list. Same practical reality as the direct API, where they’re accepted-but-ignored.
• response_format is explicitly excluded; structured output goes through OpenRouter’s structured_outputs parameter instead. Porting code that relies on response_format will need this one change.
• Tools work, tool_choice barely: tool definitions are supported, but tool_choice accepts only auto — no required, no forcing a specific function.
• Streams can’t be aborted server-side (can_abort: false). Disconnecting stops your download, not necessarily the generation — or the bill. Design retry logic accordingly.
• Web search: the provider’s native web search is supported, with per-call pricing in OpenRouter’s metadata.

And one thing that doesn’t change: orchestration billing. OpenRouter’s own model page carries Sakana’s note that orchestration tokens consumed by the system are billed as standard input/output tokens. The broker doesn’t absorb the architecture.

Data policy: read before sending anything real

Per the provider data policy on OpenRouter’s listing (August 18, 2026): the provider retains prompts (no retention duration published), does not train on them, and cannot publish them. That’s a different posture from Sakana’s direct API, where inputs and outputs are used for training by default with an opt-out. Whichever route you take, check the current policy on the respective pages before sending production data — these postures can change without notice.

Setting up a safe evaluation

1. Create an OpenRouter account and load a small amount of credit.
2. Create a dedicated key with a credit limit — your hard spending cap, and the single best reason to evaluate Fugu here rather than direct.
3. Start in our browser playground: it uses this exact route (your key, your browser, straight to OpenRouter — never our server) and prints token usage and estimated cost under every reply.
4. Graduate to the SDK with the code above, logging usage from every response.
5. If Fugu earns a production slot and you need base Fugu, the model pool, or subscriptions — that’s when a direct Sakana account (region permitting) enters the picture.

Ready to try it? The playground needs nothing but a credit-limited key. For the full direct-API picture, see the Fugu API guide; for what a serious workload will cost on either route, the calculator has the honest math.