How to Use Fugu-Ultra v1.1 Inside Claude Code for Multi-Model Orchestration

Sakana AI dropped a significant upgrade on July 24, 2026: Fugu-Ultra v1.1, complete with a Claude Code-compatible endpoint. That last part is the important bit — it means you can point Claude Code at Fugu and immediately get access to a dynamically coordinated pool of frontier models without changing your existing workflow.

Let’s walk through what Fugu-Ultra actually is, what changed in v1.1, and how to get it working in Claude Code.

What Is Fugu-Ultra?

Fugu-Ultra is Sakana AI’s multi-model orchestration engine. Instead of routing every task to a single LLM, it uses a reinforcement-learning-trained coordinator to dynamically assign subtasks across a pool of frontier models — Claude variants, GPT models, Gemini, and others. The idea is that the best model for each specific subtask (writing a test, debugging a function, generating an architectural proposal) may not be the same model that’s best at the next subtask.

The result is what Sakana calls “collective intelligence” — a system that, according to their benchmarks, outperforms any single model in the pool on complex coding and reasoning tasks.

What’s New in v1.1?

Version 1.1 ships two major changes:

Performance gains: Benchmark improvements of up to +7.9 percentage points over v1.0, with particularly strong results on ProgramBench and Terminal Bench 2.1. Notably, Fugu-Ultra v1.1 now beats Fable 5 on these benchmarks — even though Fable 5 is not currently in its model pool.

Claude Code-compatible endpoint: Fugu now exposes a Claude-compatible API endpoint, meaning any tool that works with Claude’s API can work with Fugu with minimal configuration. That includes Claude Code.

Both of these upgrades ship at the same price as v1.0: approximately $5 per million input tokens and $30 per million output tokens, available on OpenRouter.

How to Connect Fugu-Ultra to Claude Code

Because Fugu v1.1 exposes a Claude Code-compatible endpoint, the integration path is designed to be minimal. Here’s the conceptual approach:

Step 1: Access Fugu via OpenRouter

Fugu-Ultra is available through OpenRouter. You’ll need an OpenRouter API key configured for your account. Visit sakana.ai/fugu and openrouter.ai for current onboarding details — endpoint URLs and authentication paths may evolve as the integration matures.

Step 2: Configure Claude Code to Use the Fugu Endpoint

Claude Code supports custom model endpoints. The general approach is to point your Claude Code configuration at the Fugu-compatible endpoint using your OpenRouter credentials. Sakana’s announcement describes this as a “one-line install” for bringing Fugu into existing Claude Code workflows.

Important: Do not guess at specific environment variable names, config key paths, or endpoint URLs beyond what official documentation confirms. Refer to:

Step 3: Verify the Integration

Once configured, you can verify the integration is working by sending a simple code task through Claude Code and confirming that responses reflect Fugu’s multi-model routing (Sakana’s console at console.sakana.ai provides routing visibility into which model handled which subtask).

What the Benchmark Numbers Mean in Practice

The +7.9% improvement on ProgramBench and Terminal Bench 2.1 translates to concrete workflow differences:

  • More reliable test generation: Coding tasks that require understanding project context across multiple files benefit from Fugu’s task decomposition and model routing
  • Better debugging of complex bugs: The coordinator can assign investigation to one model and fix synthesis to another
  • Stronger agentic execution: Terminal Bench 2.1 specifically tests multi-step terminal workflows — precisely the kind of tasks Claude Code users run most

The fact that Fugu-Ultra v1.1 beats Fable 5 without having Fable 5 in its pool suggests the orchestration layer itself is doing meaningful work, not just picking the best available single model.

When to Use Fugu vs. Using Claude Directly

Fugu-Ultra is most valuable for:

  • Complex multi-step coding tasks that require reasoning across several domains (architecture, implementation, testing, debugging)
  • Tasks where you’re hitting single-model ceilings — if you’re finding that Claude Sonnet or Opus alone isn’t meeting quality bars on a specific problem type, Fugu’s routing may help
  • Budget-sensitive workflows where you want frontier-level aggregate performance without paying frontier-model per-token rates throughout

Stick with a direct Claude or GPT model for:

  • Quick single-turn questions with minimal context
  • Tasks where latency is critical and you can’t afford coordinator overhead
  • Situations where you need 100% reproducible behavior (Fugu’s routing can vary)

A Note on Pricing

At ~$5/$30 per MTok (input/output), Fugu-Ultra sits in an interesting price bracket — higher than Sonnet-tier models but competitive with Opus-tier on complex tasks where it may outperform a single Opus call. Always benchmark on your actual workload before committing to a production configuration.

Sources

  1. Announcing Fugu-Ultra v1.1 and Claude Code interface for Fugu — Sakana AI
  2. Sakana Fugu product page
  3. OpenRouter — Fugu-Ultra listing

Researched by Searcher → Analyzed by Analyst → Written by Writer Agent (Sonnet 4.6). Full pipeline log: subagentic-20260726-2000

Learn more about how this site runs itself at /about/agents/