Sakana Fugu in Claude Code: Setup, Models, v1.1

What Sakana AI changed in Fugu Ultra v1.1

Fugu-Ultra v1.1 is an update to the model pool that Sakana Fugu coordinates. No model was rebuilt: the company refreshed the pool with the latest frontier models.

The release also brought a Claude Code-compatible endpoint. For how Fugu routes requests and how it is billed, see Sakana Fugu Ultra is not always the best Fugu.

Fugu-Ultra v1.1 at a glance

Released

July 24, 2026

Gain over v1.0

Up to 7.9 points

Strongest results named

ProgramBench, Terminal Bench 2.1

Price

Same as v1.0

New with this release

Claude Code-compatible endpoint

Model name

fugu-ultra-v1.1

Sakana Fugu Ultra gains up to 7.9 points at the same price

Sakana AI reports that v1.1 scores up to 7.9 points higher than v1.0, at the same price. It says the best results came on ProgramBench and Terminal Bench 2.1.

The company does not list a score for each benchmark, so the size of each gain is unknown for now. On cost, a team that already pays for Fugu-Ultra v1.0 gets the newer pool at the same rate.

How Sakana AI pits Fugu Ultra against Fable 5

Sakana AI compares Fugu with Fable 5 on two pages, and the wording differs. The v1.1 blog post says Fugu now beats models like Fable 5 on complex coding and reasoning. The Fugu product page says Fugu is shoulder-to-shoulder with Fable 5 and Mythos Preview.

Both pages agree that Fable 5 is not in Fugu's pool. The product page gives the reason: Fable 5 and Mythos Preview are not publicly accessible. Both claims come from Sakana AI itself.

Claude Code setup for Sakana Fugu: two routes

Claude Code is Anthropic's coding agent that runs in the terminal. Fugu's new endpoint accepts the same kind of request that Claude Code normally sends to Anthropic.

Your operating system decides the route: the one-line installer on macOS and Ubuntu, or environment variables on Windows. The two-routes table below lists both routes side by side.

Two ways to run Sakana Fugu in Claude Code

Route	Works on	What you run
One-line installer	macOS, Ubuntu	Install, then claude-fugu
Environment variables	Any system, Windows too	Set ANTHROPIC_*, then claude

Each route has one catch. The launcher from the installer does not update itself, and the environment variables end when the terminal closes. Both routes also need an API key from the Sakana AI console, which shows the key only once, so save it before you close the dialog.

Sakana Fugu one-line Claude Code setup on macOS and Ubuntu

On a Mac or an Ubuntu machine, one command does the setup, and claude-fugu then opens Claude Code connected to Fugu. The command downloads a setup script from Sakana AI and runs it.

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

The same install adds codex-fugu for people who use the Codex CLI. The command pipes a script from the internet straight into your shell, so if your company reviews scripts before they run, download it first and read it. If the install stops partway, the manual route gets you to the same setup.

sakutto.ai downloaded the installer on November 6, 2026 and read it without running it: the 123-line bash script asks whether to set up Codex or Claude Code, and when it cannot show that menu it installs Codex only. It does four things.

What the Fugu installer does (script fetched November 6, 2026 and reviewed by sakutto.ai)

Step	What happens
Download	git clones the SakanaAI/fugu repository into ~/.fugu (the FUGU_HOME variable changes the location)
Choose a target	A menu offers "1) Codex", "2) Claude Code" and "3) Both"; the --claude, --codex or --both argument, or the FUGU_INSTALL_TARGET variable, does the same
No terminal for the menu	With no argument and no variable, it installs Codex only
Later runs	If a needed file is missing it resets ~/.fugu to the latest version, discarding local changes

To be sure you get the Claude Code setup, pass the argument: curl -fsSL https://sakana.ai/fugu/install | bash -s -- --claude. If you have edited files inside ~/.fugu, copy them elsewhere before running the installer again.

Claude Code on Windows with environment variables

An environment variable is a named setting that a program reads when it starts. Claude Code checks a group of variables that start with ANTHROPIC_ to decide where to send requests. The installer supports only macOS and Ubuntu, so on Windows you set these variables yourself and then run claude.

Sakana AI's example uses the export command from macOS and Linux shells. In Windows PowerShell, the same setting is written as $env:ANTHROPIC_BASE_URL = "https://api.sakana.ai", and the other lines follow that pattern.

export ANTHROPIC_BASE_URL="https://api.sakana.ai"
export ANTHROPIC_AUTH_TOKEN="fish_..."   # your Sakana key
export ANTHROPIC_DEFAULT_OPUS_MODEL="fugu-max[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL="fugu[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="fugu[1m]"
export ANTHROPIC_DEFAULT_FABLE_MODEL="fugu-ultra[1m]"
export CLAUDE_CODE_SUBAGENT_MODEL="fugu[1m]"
claude

These settings last only for the terminal window where you typed them. Open a new window and Claude Code goes back to Anthropic. That suits a trial run, since the change ends with the window.

Anthropic base URL and auth token values for Sakana Fugu

Set ANTHROPIC_BASE_URL to https://api.sakana.ai and put your Sakana key in ANTHROPIC_AUTH_TOKEN. The first is the address Claude Code sends requests to, and the second proves the requests are yours.

The variable name matters here. Sakana AI asks for ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY. Claude Code sends that variable as a bearer token.

Claude Code model mapping to Fugu Ultra and effort levels

Claude Code sorts its models into tiers named after Anthropic models: Opus is the strongest, Sonnet is the everyday tier, and Haiku is the light tier. Under Fugu, Opus points to fugu-max[1m], Sonnet and Haiku to fugu[1m], and Fable to fugu-ultra[1m]. The tier mapping table shows the full mapping in Sakana AI's current setup guide.

Claude Code tiers and the Fugu model behind each

Claude Code tier	Fugu model
Opus	fugu-max[1m]
Sonnet	fugu[1m]
Haiku	fugu[1m]
Fable	fugu-ultra[1m]
Subagents	fugu[1m]

Sessions start on fugu-max[1m], the Opus tier, at high effort. Subagents get their model from CLAUDE_CODE_SUBAGENT_MODEL. The setup guide does not say what the [1m] ending means, so copy the names exactly as written.

The bare name fugu-ultra now follows Sakana AI's newest Ultra release. To run v1.1 in particular, add ANTHROPIC_CUSTOM_MODEL_OPTION="fugu-ultra-v1.1[1m]"; a fixed v1.1 row then appears in the /model menu. A tier can point at fugu-cyber[1m] only if Sakana AI has approved your key for it.

Access to fugu-cyber is covered in Sakana Fugu Cyber: 86.9% on CyberGym, How to Apply. Effort is how hard the model thinks before it answers, and fugu-ultra-v1.1 offers 3 levels: high, xhigh and max. The other Fugu models have 2 levels (high and xhigh) and treat max as xhigh.

Three Claude Code limitations Sakana AI documents

Claude Code is closed source, so Sakana AI cannot change how it looks or behaves from the inside. Fugu works only through the ANTHROPIC_* variables and the company's own server. The three documented limits are label mismatches on screen, a launcher that does not update itself, and a provider pool that includes every supported provider by default.

Claude Code model picker and effort slider mismatches

Some parts of the Claude Code screen still assume Anthropic models. Sakana AI says these are cosmetic only: the model you talk to, its answers, tool use and subagents all work as you set them up. The three mismatches are these:

• Effort slider: Claude Code shows 6 stops, from low to ultracode. Fugu has 2 real levels and splits the stops at the line between high and xhigh. Stops on the same side give the same result.
• Model picker: With a recent version of Claude Code and claude-fugu, the menu lists one row per Fugu model. With an older version or the manual setup, Sonnet and Haiku both point to fugu[1m], so 2 identical rows appear. Their descriptions name the Anthropic tier, not the Fugu model.
• Session header: The billing and default-tier labels describe API-token use.

Claude Code changes on its own schedule, so new mismatches of this kind may appear later. None of the three changes what Fugu returns.

Sakana Fugu launcher skips Claude Code setup updates

The claude-fugu launcher does not update itself. When Sakana AI adds a new Fugu model for Claude Code, a launcher installed earlier will not know about it.

To get the change, run the install command again, or apply the update by hand from the Get Started page. On the environment-variable route, you update the model names in your own variables when Sakana AI changes the mapping.

Sakana Fugu routing pool you control in Sakana AI console

By default, fugu can send a request to any provider it supports, and you limit that list on the API key. You do not pick the provider for each request. If your company has rules about where code may go, this is the setting to look at first.

To leave a provider out, open the API key in the Sakana AI console. Turn on "Fugu custom model pool" when you create or edit the key, and keep only the providers you allow. With the setting off, Fugu uses its full default pool.

Try Sakana Fugu in Claude Code on one real task

Fugu-Ultra v1.1 brings a newer model pool at the v1.0 price, and Claude Code can now use it with little setup. On macOS or Ubuntu, start with the one-line installer and claude-fugu. On Windows, set the ANTHROPIC_* variables in one terminal window and run claude.

Sakana AI has not published per-benchmark scores for v1.1. The most direct check is to give Fugu one task from your own repository and compare the result with the model you use today. If you want v1.1 itself rather than the newest Ultra, add the pinned v1.1 option first.

If you also set up Fugu in the Codex CLI, the manual steps ask you to save a model list as a JSON file. To check that file for syntax errors before you save it, try the JSON formatter and validator, which shows the line where the syntax breaks.