Sakana AI · 2026-09-11 · major
Fugu Max and Fugu Ultra v2 — Sakana's router splits into cheap and strong
Sakana AI released Fugu Max and Fugu Ultra v2, two versions of its orchestrator that routes each task to other models. Fugu Max costs $2 and $6 per million tokens; Fugu Ultra v2 scores 48.3 on Chartography against Opus 5's 27.3.
Sakana AI splits its Fugu orchestrator into a cheap tier and a high-capability tier, both behind one OpenAI-compatible API.
Quick facts
| Maker | Sakana AI |
|---|---|
| Models | Fugu Max, Fugu Ultra v2 |
| Type | Orchestrator that routes to other models |
| Availability | OpenAI-compatible API, live now |
| Also on | OpenRouter, Vercel AI Gateway |
| Ultra v2 training cutoff | 28 August 2026 |
Benchmarks
Pricing
| Fugu Max — input | $2.00 / 1M tokens |
|---|---|
| Fugu Max — output | $6.00 / 1M tokens |
| Fugu Ultra v2 — input · $10.00 above 272K context | $5.00 / 1M tokens |
| Fugu Ultra v2 — output · $45.00 above 272K context | $30.00 / 1M tokens |
What is it?
Fugu Max is the cheap end of Sakana AI's orchestrator family — it sends each task to the leanest model that can solve it, at $2 per million input tokens and $6 per million output tokens. Fugu Ultra v2 takes the other end, aiming for the best result on hard multi-step work such as reasoning, autonomous research and software development. Neither is a single foundation model.
How does it work?
The orchestration engine reads a query, builds an agent scaffold for it, and routes the work across a pool of open-weight and specialist models, including NVIDIA Nemotron models. Sakana says Fugu Ultra v2 reaches its scores without Fable 5, Fable 5.1 or GPT-6 Astra in that pool. The technical report describes training the Fugu routers with large-scale fine-tuning, evolutionary algorithms and reinforcement learning.
Why does it matter?
Two price points now sit on the same API, so a team can trade cost against capability with a one-line parameter change instead of a rewrite. Sakana puts Fugu Max's output price 40–60% below Sonnet 5, GPT 5.6 Terra and Kimi K3, and reports best overall scores for it on six benchmarks, among them Terminal Bench 2.1 and GPQAD.
Who is it for?
teams paying per token for agent work
Frequently asked questions
- How much cheaper is Fugu Max than Fugu Ultra v2?
- Fugu Max runs at $2 per million input tokens and $6 per million output tokens at any context length. Fugu Ultra v2 is $5 input and $30 output per million tokens, rising to $10 and $45 above 272K context. So Fugu Max costs a fifth as much on output for the same API.
- Which models does Fugu Ultra v2 route to?
- Fugu Ultra v2 routes across a pool of open-weight and specialist models, including NVIDIA Nemotron models. Sakana AI notes that the pool does not contain Fable 5, Fable 5.1 or GPT-6 Astra, so the reported scores come from orchestrating smaller and cheaper models rather than from calling a frontier model behind the scenes.
- Do I have to change my code to move from Fugu to Fugu Max?
- No. Sakana AI says existing Fugu users switch with a single-line parameter change, because all the Fugu variants sit behind the same OpenAI-compatible API endpoint. You keep your client library, your API key and your request shape, and only the model parameter changes.
- Can I use Sakana Fugu outside Sakana's own API?
- Yes. Alongside Sakana AI's own OpenAI-compatible endpoint and the console at console.sakana.ai, the Fugu models are offered through OpenRouter, the Vercel AI Gateway and other providers. Sakana also sells subscription plans at $20, $100 and $200 a month instead of pay-as-you-go tokens.
Try it
Sign in at console.sakana.ai and switch the model parameter to Fugu Max.