AI/TLDR

Sakana Fugu

An orchestrator that routes each task to the leanest model that can solve it

Gateways & RoutingCommercial
Latest
fugu-ultra-v2.0
Updated
11 Sep 2026
Coverage
1 story

What's new

fugu-ultra-v2.011 Sep 2026

Sakana AI split the Fugu orchestrator into a cheap tier and a high-capability tier: Fugu Max at $2/$6 per million input/output tokens, and Fugu Ultra v2 at $5/$30 rising to $10/$45 above 272K context. Both are served from the existing OpenAI-compatible endpoint.

Latest news

Overview

Fugu is Sakana AI's orchestration layer, not a foundation model. Given a query it builds an agent scaffold for that task and routes the work across a pool of open-weight and specialist models — NVIDIA Nemotron models among them — picking which model handles which step rather than sending everything to one frontier endpoint. Sakana's technical report describes training the Fugu routers with large-scale fine-tuning, evolutionary algorithms and reinforcement learning.

Four variants sit behind the same API. Fugu is the balanced default; Fugu Ultra optimises for performance; Fugu Max optimises for cost-performance; Fugu Cyber is specialised for cybersecurity. Because they share one OpenAI-compatible endpoint, moving between them is a change to the model parameter — Sakana states no SDK migration is required, so the client library, API key and request shape stay as they are.

Pricing differs sharply across the tiers, which is the point of having them. Fugu Max is $2 per million input tokens and $6 per million output, with cached input at $0.25 and web functions at $0.007 per call. Fugu Ultra (fugu-ultra-v2.0) is $5 input and $30 output per million, rising to $10 and $45 above 272K context, with cached input at $0.50 (or $1.00 above 272K). Plain Fugu is charged at the standard rate of the underlying models, billed at a single rate based on the top-tier model involved when multiple agents are active; Fugu Cyber is sales-contact only. Subscriptions are offered at $20, $100 (10x usage) and $200 (20x usage) per month, all tiers including all four models. Alongside Sakana's own endpoint and console, the models are served through OpenRouter, Vercel, Models.dev, Creao and Merge.

What it does

  • Per-query agent scaffolding: the orchestration engine reads a request and routes its steps across a pool of open-weight and specialist models
  • Four variants — Fugu, Fugu Ultra, Fugu Max and Fugu Cyber — behind one OpenAI-compatible API, switched by the model parameter
  • Distinct price points per variant, from Fugu Max at $2/$6 per million input/output tokens to Fugu Ultra at $5/$30
  • Cached-input pricing ($0.25 on Fugu Max, $0.50 on Fugu Ultra) and per-call billing for web functions at $0.007
  • Long-context tier: Fugu Ultra prices rise to $10 input and $45 output per million above 272K context
  • Available through third-party platforms as well as Sakana's own endpoint — OpenRouter, Vercel, Models.dev, Creao and Merge
  • Subscription plans at $20, $100 and $200 per month as an alternative to pay-as-you-go tokens

Getting started

Fugu is a hosted service, so there is nothing to install — you point an existing OpenAI-compatible client at Sakana's endpoint and set the model parameter. Sign in at console.sakana.ai to get an API key.

Get an API key

Create an account and issue a key from the Sakana console.

texttext
https://console.sakana.ai

Choose a variant by price point

Fugu Max is the cost-performance tier, Fugu Ultra the performance tier, plain Fugu the balanced default, and Fugu Cyber the cybersecurity specialisation (sales contact required). Check the current rates on the Models & Pricing page before committing.

texttext
https://sakana.ai/fugu/

Switch the model parameter

Because the endpoint is OpenAI-compatible, Sakana says no SDK migration is required — existing users move between Fugu variants with a single-line parameter change.

Or route through a third-party gateway

The Fugu models are also offered via OpenRouter, Vercel, Models.dev, Creao and Merge, which is the simpler path if you already consolidate providers behind one of those.

Commands and code are distilled from the project's own documentation — always check the official docs for the latest.

When to use it

  • Trade cost against capability on agent workloads by moving between Fugu Max and Fugu Ultra with a parameter change rather than a rewrite
  • Get orchestrated multi-model results without building and maintaining your own routing layer
  • Cut spend on long-running agent work where most steps do not need a frontier model
  • Evaluate a routing-based orchestrator against single-model baselines through an OpenAI-compatible endpoint you already support

How Sakana Fugu compares

Sakana Fugu alongside other open-source gateways & routing tools AI/TLDR tracks, ranked by GitHub stars.

ToolStarsWhat it does
OmniRoute★ 64.8kA local-first AI gateway that fronts hundreds of model providers with one OpenAI-compatible endpoint, quota-aware auto-fallback, many routing strategies and prompt compression.
LiteLLM★ 58.5kA Python SDK and proxy server that gives one OpenAI-compatible API to 100+ LLM providers, with cost tracking, budgets, fallbacks, rate limiting, and an admin UI.
Free Claude Code★ 54.5kA local proxy that points Claude Code, Codex, Pi, OpenCode, Cline and other agent CLIs at 50 free, paid, subscription or local model providers, with automatic failover and an admin UI.
Kong Gateway★ 44.1kA Lua/NGINX API gateway whose AI Gateway plugins put one universal endpoint in front of OpenAI, Anthropic, Gemini, Bedrock and other providers, with semantic routing, caching and MCP traffic governance.
Apache APISIX★ 17.1kA cloud-native API gateway whose AI plugins add multi-provider LLM proxying, load balancing, retries and fallbacks, token-based rate limiting, and content moderation.
Portkey AI Gateway★ 13kAn LLM gateway that routes calls to 100+ providers through one API and adds logging, tracing, caching, and fallbacks for production AI traffic.
Higress★ 9.4kAn AI-native API gateway built on Istio and Envoy that proxies and governs traffic to many LLM providers, with token rate limiting, caching, and MCP server hosting.
Sakana FuguAn orchestrator that routes each task to the leanest model that can solve it