Running Sakana Fugu in Claude Code: Setup Steps and What Changed in v1.1

What changed in Fugu-Ultra v1.1

Sakana Fugu is a multi-agent system from the Japanese AI company Sakana AI that automatically orchestrates multiple frontier models and exposes them as a single API. Rather than competing on the intelligence of one model, it is designed to dynamically assemble whichever models suit the task at hand. We cover the underlying mechanism and pricing in our article on Sakana Fugu.

v1.1 is an update to that orchestration engine. No model was rebuilt; the change is that the pool being orchestrated was refreshed with the latest frontier models.

Up to +7.9 points, at the same price

The only figure published is "gains of up to 7.9 points over v1.0." ProgramBench and Terminal Bench 2.1 are named as showing particularly strong results, but no individual scores are given. There is currently no way to check what went from what to what. If you want a numeric comparison, you are waiting on Sakana AI to publish more.

The cost side is unambiguous. The improvement ships at the same price as v1.0. From a user's perspective, the bill stays put while the model lineup underneath gets newer. That is the orchestration-engine design paying off directly.

"Shoulder-to-shoulder" and "beating" coexist

One more thing worth noticing. The product page still describes Fugu as "shoulder-to-shoulder" with Claude Fable 5 and Mythos Preview. The v1.1 announcement, meanwhile, says Fugu is "beating leading models like Fable 5 in complex coding and reasoning tasks." Two pages from the same company do not agree on how strong the claim against Fable 5 is.

The stronger claim also carries a condition: Fugu beats Fable 5 without having Fable in its agent pool. That is not a deliberate exclusion, though. The product page states the reason directly — Fable 5 and Mythos Preview are not in the pool because they are not publicly accessible. Orchestrating only what is reachable in order to match models above it is the design, and it places Fugu on different ground from a single-model capability race.

All of this remains the developer's own claim, with third-party verification still to come. As long as individual scores stay unpublished, there is nowhere to go beyond "this is what they say."

Using Fugu from Claude Code

The Claude Code-compatible endpoint shipped alongside v1.1. Claude Code is Anthropic's terminal coding agent, but pointing it somewhere else lets you call Fugu's orchestration directly. Sakana AI frames it as this: rather than relying on a single model to write, debug, and execute your code, you can orchestrate a diverse pool of state-of-the-art models straight from your terminal.

There are two routes: a one-liner on supported systems, and manual environment variables everywhere else.

Starting with the bundled installer (Ubuntu / macOS)

The documented path is fetching an install script and running it. Afterwards, claude-fugu launches Claude Code already pointed at Fugu (Codex CLI users get codex-fugu).

curl -fsSL https://sakana.ai/fugu/install | bash

This one-liner supports Ubuntu and macOS only. On Windows, or if the install does not complete, you move to the manual setup below. Since the form of it pipes a fetched script straight into a shell, reading the contents before running is the safer habit — downloading it and looking first changes nothing about the result.

Setting environment variables by hand (the Windows route)

You can skip the launcher entirely. Claude Code switches its endpoint through a set of ANTHROPIC_* environment variables, so exporting them and running claude points it at Fugu.

export ANTHROPIC_BASE_URL="https://api.sakana.ai"
export ANTHROPIC_AUTH_TOKEN="fish_..."
export ANTHROPIC_DEFAULT_OPUS_MODEL="fugu-ultra[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL="fugu[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="fugu[1m]"
export ANTHROPIC_DEFAULT_FABLE_MODEL="fugu-cyber[1m]"  # optional (access-gated)
export CLAUDE_CODE_SUBAGENT_MODEL="fugu[1m]"
claude

Two things trip people up. First, the variable is ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY — it is treated as a bearer token, not an API key. Second, this setup lasts only for the current shell session. For everyday use, put it in your shell configuration or wrap it in a script.

The [1m] suffix on the model names appears verbatim in the official examples. Its meaning is not explained on the Get Started page, so copying it exactly as written is the reliable move.

If you also use the Codex CLI, there is a step where the model catalog is saved to ~/.codex/fugu.json, which means editing JSON directly. A single character out of place will stop it from starting, so running the file through a formatter and validator before pasting it saves grief.

Free ToolJSON Formatter & ValidatorPretty-print or minify JSON data. Catch syntax errors instantly with line numbers and tree view.Try it now →

Model mapping and effort levels

Claude Code assumes Anthropic's Opus / Sonnet / Haiku tiers, so you specify which Fugu model each one resolves to. The mapping is as follows.

Reasoning effort differs too. fugu-ultra-v1.1 supports three levels — high, xhigh, and max, where max is a distinct maximum for the hardest problems. By contrast, fugu, fugu-ultra-v1.0, and fugu-cyber support only high and xhigh; max is accepted for compatibility and maps to xhigh. For fugu-cyber specifically, see our article on Fugu-Cyber.

Three limitations to know before you start

Connecting successfully and using it without friction are two different things. There are three places where Sakana AI itself flags a mismatch.

The display mismatches are "cosmetic only"

Because Claude Code is closed source, Sakana AI can only drive it through the documented ANTHROPIC_* variables and its own server-side gateway. As a result, parts of the interface keep Anthropic's assumptions. Sakana AI states these are "cosmetic only," and that the model you talk to, its responses, streaming, tool use, and subagents all behave exactly as configured.

The three specific mismatches:

• Effort slider: Claude Code shows six stops (low through ultracode), but Fugu has two real levels. It maps them at the high/xhigh boundary, so stops within a group are equivalent
• Model picker: Sonnet and Haiku both resolve to fugu[1m], so two identical rows appear. Without the launcher's curated labels, descriptions name the Anthropic tier rather than the Fugu model
• Session header: billing and default-tier labels reflect API-token usage

Because Claude Code evolves independently, new cosmetic mismatches may surface over time.

The launcher does not auto-update

There is no auto-update mechanism in the claude-fugu launcher. When Claude Code support is updated — new Fugu models, for instance — your local launcher stays where it was. Sakana AI's instruction is to check the Get Started page and apply the update manually. Installing it once is not the end; periodically checking back is assumed.

If you went the environment-variable route, this is easier, since changing a model name is all it takes.

You can pick the routing targets, but not see inside

The fugu model routes across all supported providers by default. To opt out of a provider, enable "Fugu custom model pool" while creating or editing the API key and leave only the providers you want. Left off, the full default pool is used.

Being able to filter by provider means that, by default, the user is not choosing which provider a request goes to. When data handling has to be controlled tightly, that setting is worth checking first.

Wrapping up: rather than wait for numbers, try it on your own task

Fugu-Ultra v1.1 raises performance by swapping the orchestrated pool instead of rebuilding a model, and holds the price flat. The intent to win on different ground from a single-model race shows plainly in the claim of beating Fable 5 without Fable in the pool. Individual scores stay unpublished, though, and verification is left to third parties.

Rather than wait for numbers, the practical gain here is the lowered barrier to using Fugu from Claude Code. Throwing one real task at it and seeing the difference against whatever model you use today is faster than reading about it. On Ubuntu or macOS the one-liner is the starting point; on Windows, the manual environment variables are.

Free ToolJSON Formatter & ValidatorPretty-print or minify JSON data. Catch syntax errors instantly with line numbers and tree view.Try it now →