Sakana Fugu's Claude Code-compatible interface, released on July 24, 2026 alongside Fugu-Ultra v1.1, lets you replace Claude Code's single-model backend with a multi-agent orchestrator that dynamically coordinates a pool of frontier models — all from the same terminal workflow you already use. The setup takes one install command and one launch command, costs $20/month on the cheapest subscription or $5/$30 per million tokens on pay-as-you-go, and works on macOS and Ubuntu. The catch: you cannot see which underlying models handle each request, the launcher does not auto-update, and it is not available in the EU or EEA. For developers already using Claude Code who want multi-agent orchestration without switching tools, this is the fastest path — but it is not the right choice for every task.
Last verified: 2026-08-02 · One-line install +
claude-fugulauncher · Opus tier maps to Fugu Ultra, Sonnet/Haiku to Fugu · macOS and Ubuntu only · $20-$200/month subscription or pay-as-you-go · EU/EEA restricted · Vendor-reported benchmarks — test on your own workloads
What Is the Sakana Fugu Claude Code Interface?
The Sakana Fugu Claude Code interface is a compatibility layer that lets Claude Code — Anthropic's terminal-based AI coding agent — send requests to Sakana Fugu's API instead of Anthropic's models. Sakana AI announced it on July 24, 2026, alongside the Fugu-Ultra v1.1 update. Claude Code remains the agent harness (it handles file I/O, tool calls, and the terminal interface), while Fugu becomes the reasoning backend that decides how to decompose tasks, delegate to specialized models, verify their work, and synthesize results. Source: Sakana AI Fugu-Ultra v1.1 announcement
This is not a new Claude Code feature from Anthropic. It is a Sakana-side endpoint that speaks Anthropic's API protocol. You provide a Sakana API key instead of an Anthropic key, and Claude Code routes its requests to api.sakana.ai instead of Anthropic's servers. The distinction matters for billing, data privacy, and model visibility: you are now in Sakana's ecosystem, not Anthropic's. Source: Sakana console get-started
How Does Sakana Fugu Work Underneath Claude Code?
Fugu is itself a small language model trained specifically to orchestrate other models. When you send a request through Claude Code, the Fugu orchestrator decides whether to answer directly or assemble a team of expert models from its internal pool. It handles model selection, task delegation, verification, and synthesis internally — so from Claude Code's perspective, it looks like calling one model. The architecture is built on two ICLR 2026 papers: TRINITY (an evolutionarily optimized LLM coordinator that assigns Thinker/Worker/Verifier roles) and the Conductor (learning to orchestrate agents in natural language). Source: Sakana Fugu technical report (arXiv)
Think of it this way: normally Claude Code talks to one model (say, Claude Sonnet or Opus). With Fugu, Claude Code talks to a conductor model that quietly assembles a team of frontier models — potentially including GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro, and others — splits the work, verifies it, and returns one unified answer. You never pick which model does what. Fugu picks for you, task by task.
How Do You Install Fugu in Claude Code?
The installation takes two commands on macOS or Ubuntu. First, create an API key in the Sakana console — the key is displayed only once, so store it securely. Then run the one-line installer and launch:
Step 1: Install the Fugu launcher
curl -fsSL https://sakana.ai/fugu/install | bash
Step 2: Launch Claude Code with Fugu
claude-fugu
That is it. The installer clones the SakanaAI/fugu repository, deploys the configuration, stores your API key, and sets the environment variables Claude Code reads. Every message you send from this point goes to Fugu's orchestrated pool instead of a single Anthropic model. Source: SakanaAI/fugu GitHub repository
The one-line installer officially supports macOS and Ubuntu. On Windows, use WSL Ubuntu or set up the configuration manually.
How Do You Set Up Fugu in Claude Code Manually?
If the installer does not work in your environment, or you prefer to control the configuration yourself, Claude Code reads its backend from a set of ANTHROPIC_* environment variables. No launcher is required — you export them in your shell and run claude directly:
export ANTHROPIC_BASE_URL="https://api.sakana.ai"
export ANTHROPIC_AUTH_TOKEN=*** # your Sakana key
export ANTHROPIC_DEFAULT_OPUS_MODEL="fugu-ultra[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL="fugu[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="fugu[1m]"
export CLAUDE_CODE_SUBAGENT_MODEL="fugu[1m]"
claude
Key details: use ANTHROPIC_AUTH_TOKEN (a bearer token), not ANTHROPIC_API_KEY. The Opus tier maps to fugu-ultra[1m], Sonnet and Haiku both map to fugu[1m], and subagents use fugu[1m] by default. This configuration lasts for the current shell session — add it to your shell profile for persistence. Source: Sakana console get-started
What Changes Inside Claude Code When You Use Fugu?
The interface looks familiar, but the model semantics are different. Claude Code's built-in Anthropic assumptions remain visible even though Fugu is powering the responses.
| Claude Code surface | What Fugu does | Cosmetic? |
|---|---|---|
| Opus selection | Routes to fugu-ultra[1m] |
No — this is the real backend |
| Sonnet selection | Routes to fugu[1m] |
No |
| Haiku selection | Also routes to fugu[1m] |
Yes — two identical-looking rows in the model picker |
| Subagents | Use fugu[1m] by default |
No |
| Effort slider (6 stops) | Collapses into Fugu's high or xhigh groups |
Yes — stops within a group are equivalent |
| Model picker labels | May show Anthropic-oriented names | Yes — cosmetic mismatch |
| Session header | Billing labels reflect Claude Code assumptions | Yes |
Sakana confirms these mismatches are cosmetic: model responses, streaming, tool calls, and subagents all follow the configured Fugu backend. Two rows in the model picker reading the same is normal — it is because both Sonnet and Haiku resolve to fugu[1m]. Source: Sakana console get-started
What Is the Difference Between Fugu and Fugu Ultra in Claude Code?
Fugu is the lower-latency default optimized for everyday coding, code review, chat, and interactive work. Fugu Ultra coordinates a deeper pool of expert models and is tuned for maximum answer quality on hard, multi-step problems. Both are accessible through the same Claude Code interface, and you switch between them by selecting the Opus tier (Fugu Ultra) or the Sonnet/Haiku tier (Fugu) in Claude Code's model picker.
| Feature | Fugu | Fugu Ultra |
|---|---|---|
| Optimized for | Latency, everyday work | Quality, complex multi-step tasks |
| Agent pool | Configurable (opt-out supported) | Fixed (full pool required) |
| SWE-Bench Pro | 59.0% | 73.7% |
| Terminal-Bench 2.1 | 80.2% | 82.1% |
| LiveCodeBench v6 | 90.3% | 92.0% |
| Pricing (pay-as-you-go) | Rate of top model used | $5 input / $30 output per 1M tokens |
| Best for | Coding, chat, code review | Research, multi-step engineering, security |
Source: Sakana Fugu technical report (arXiv), evals.report
One important difference for privacy and compliance: Fugu lets you opt out of specific model providers from the console's settings menu (useful if you have data residency requirements or vendor restrictions). Fugu Ultra's pool is fixed — you cannot exclude providers because it relies on the full agent set to achieve its benchmark performance.
Fugu-Ultra v1.1, released July 24, 2026, upgraded the orchestration pool with the latest frontier models and improved scores across every benchmark Sakana tracks, with gains of up to 7.9 points over v1.0. The strongest improvements were on ProgramBench (72.4% to 79.8%) and Terminal Bench 2.1 (82.1% to 90.0%). Pricing remained unchanged. Source: Sakana AI Fugu-Ultra v1.1 announcement
How Much Does Sakana Fugu Cost?
Sakana Fugu offers two billing models: subscription plans for everyday use and pay-as-you-go API rates for heavier workloads. Both include access to Fugu and Fugu Ultra.
Subscription plans:
| Plan | Price | Usage | Best for |
|---|---|---|---|
| Standard | $20/month | Baseline | Occasional API calls, personal experiments |
| Pro | $100/month | 10x Standard | Regular coding and research sessions |
| Max | $200/month | 20x Standard | Heavy, long-running workloads |
Pay-as-you-go token pricing (Fugu Ultra):
| Cost component | Standard context (up to 272K) | Extended context (above 272K) |
|---|---|---|
| Input tokens | $5.00 per 1M | $10.00 per 1M |
| Output tokens | $30.00 per 1M | $45.00 per 1M |
| Cached input | $0.50 per 1M | $1.00 per 1M |
Standard Fugu (non-Ultra) pricing on the token plan works differently: you are billed at the rate of whichever top-tier underlying model handles the request, and Sakana does not stack per-agent fees when multiple models coordinate. Source: Sakana console pricing
A note on actual cost: Fugu's orchestration consumes additional tokens (planning, verification, synthesis) that single-model APIs do not. The per-request API response separates these so you can see the full picture, but your effective cost per task can be higher than the headline token rate suggests. Monitor the usage report before handing Fugu a broad refactor or an unbounded retry loop.
When Should You Use Fugu vs Fugu Ultra in Claude Code?
Start with plain Fugu for everyday work and switch to Fugu Ultra only when the task is genuinely hard. In Claude Code, you switch by selecting the model tier — Opus for Fugu Ultra, Sonnet/Haiku for Fugu. Here is a decision framework:
Use plain Fugu when:
- You are writing or editing individual files and components
- You need fast, interactive responses during a coding session
- The task is a known pattern (add an API endpoint, fix a common bug, write a test)
- You want to minimize cost and latency
- You need to opt out of specific model providers for compliance
Use Fugu Ultra when:
- The task requires multi-step reasoning across many files
- You need debugging that traces through complex flows and dependencies
- You are doing code review on unfamiliar or large codebases
- You want a Thinker/Worker/Verifier team to outperform a single model
- You are doing research-level analysis or security work
- Your task benefits from independent exploration and verification
The effort slider in Claude Code maps to Fugu's reasoning levels: the six UI stops collapse into high and xhigh. Bumping it to max (which maps to xhigh) makes Fugu think much deeper but also takes longer — set a higher timeout for complex Fugu Ultra jobs. This is the setting most people get wrong on day one: they crank it to max for everything and wonder why it feels slow.
What Are the Real Benchmark Numbers for Fugu Ultra?
Fugu Ultra's benchmark scores are vendor-reported, drawn from Sakana's technical report. They have not been independently reproduced. Treat them as a hypothesis to test on your own workload, not a guaranteed outcome.
| Benchmark | Fugu Ultra | Fugu | Claude Opus 4.8 | GPT-5.5 |
|---|---|---|---|---|
| SWE-Bench Pro | 73.7% | 59.0% | 69.2% | 58.6% |
| Terminal-Bench 2.1 | 82.1% | 80.2% | 74.6% | 78.2% |
| LiveCodeBench v6 | 92.0% | 90.3% | 90.3% | 90.7% |
| LiveCodeBench Pro | 90.8% | 87.8% | 84.8% | 88.4% |
| GPQA Diamond | 95.5% | 95.5% | 92.0% | 93.6% |
| Humanity's Last Exam | 50.0% | 47.2% | 49.8% | 41.4% |
| CharXiv Reasoning | 86.6% | 85.1% | 84.2% | 84.1% |
| SciCode | 58.7% | 60.1% | 53.5% | 56.1% |
Source: Sakana Fugu technical report (arXiv), verified via evals.report
Fugu-Ultra v1.1 improved on v1.0 with gains of up to 7.9 points, with the strongest improvements on ProgramBench and Terminal Bench 2.1 (82.1% to 90.0%). Sakana also reports Fugu-Ultra beating Claude Fable 5 on several benchmarks, though Fable 5 is not part of Fugu's underlying model pool — making the comparison a genuine result against a rival system, not an apples-to-apples internal comparison. Source: Sakana AI Fugu-Ultra v1.1 announcement
What Should You Watch Out For?
Five pitfalls catch people on day one:
The launcher does not auto-update. When Sakana adds newer models or changes mappings, you must re-run the install command or manually update the configuration. Check the Sakana get-started page periodically.
The effort slider caveat. Cranking effort to max for every task adds latency without proportional quality gains on simple work. Use max (maps to Fugu's
xhigh) only for genuinely hard problems, and set a higher timeout so Fugu Ultra does not time out mid-thought.You cannot see which models handled your request. Fugu's routing is proprietary. You know you are using Fugu, but not which specific underlying models answered or how they communicated. For regulated teams requiring auditable routing, this is a hard limitation.
EU/EEA availability. Sakana Fugu is not currently available in the EU or EEA while Sakana works toward GDPR compliance. If that is where you are, you cannot use it yet.
Two identical model rows are normal. In the model picker, Sonnet and Haiku both resolve to
fugu[1m], so two rows look the same. This is not a bug — it is a cosmetic mismatch from Claude Code's closed-source interface.
How Do You Evaluate Whether Fugu Is Worth It for Your Work?
Run the same commit-sized task through your current single-model setup and through Fugu, then compare. Avoid judging from a greenfield demo — use a representative bug with awkward repository context, a concrete acceptance test, and a clean git diff.
- Did Fugu identify the right files before editing?
- Did its patch pass the repository's existing tests?
- Did it catch a failure the single-model run missed?
- How long did the first useful answer take?
- What did the request cost, including output and retries?
- Can you explain which providers were permitted to see your code?
For teams with compliance requirements: standard Fugu lets you opt out of specific providers from the console, but Fugu Ultra requires the full fixed pool. You can disable training on your data in the Sakana console — do this before testing proprietary repositories.
What This Means for You
If you already use Claude Code and want multi-agent orchestration without switching tools, the Fugu integration is the fastest way to get there — install in one line, launch in one command, keep your workflow. Start with plain Fugu for interactive work at $20/month, reserve Fugu Ultra for genuinely hard multi-step tasks, and always test on your own codebase before trusting vendor benchmark numbers. For a deeper look at the underlying orchestration architecture, see our multi-agent orchestration guide. For tips on managing your Claude Code context and tokens efficiently alongside this integration, read our free Claude Code token-reduction tools. If you are weighing multiple agent harnesses, our AI agent operating system decision framework covers how to pick the right one. And for stopping AI-generated code from accumulating technical debt — a real risk when multiple models contribute — the Guide-Verify-Solve framework is a must-read.
FAQ
Q: Do I need an Anthropic API key to use Fugu in Claude Code?
A: No. You need a Sakana API key, which you pass as ANTHROPIC_AUTH_TOKEN (a bearer token, not ANTHROPIC_API_KEY). Create one in the Sakana console. Claude Code routes requests to Fugu's API instead of Anthropic's servers.
Q: Does the claude-fugu launcher auto-update?
A: No. When Sakana adds new models or changes mappings, you must re-run the installer (curl -fsSL https://sakana.ai/fugu/install | bash) or manually update the configuration. Check the Sakana get-started page periodically for changes.
Q: Is Fugu available in the EU or EEA? A: No. Sakana Fugu is not currently available in the EU or EEA while Sakana works toward GDPR compliance. This is a product constraint, not a setup issue — if you are in those regions, you cannot use Fugu yet.
Q: Which model do I actually talk to when I select "Opus" in Claude Code?
A: Selecting Opus routes your request to fugu-ultra[1m], Sakana's deep-pool orchestration model. Selecting Sonnet or Haiku routes to fugu[1m], the lower-latency everyday model. Subagents use fugu[1m] by default.
Q: Can I control which underlying models Fugu uses? A: Partly. With standard Fugu, you can opt out of specific providers from the Sakana console's settings menu for data privacy or compliance. Fugu Ultra requires the full fixed pool — you cannot exclude providers because it relies on the complete agent set for benchmark performance.
Q: Is Fugu Ultra v1.1 actually better than Claude Fable 5? A: On several of Sakana's own benchmarks, Fugu-Ultra v1.1 outperforms Claude Fable 5 (for example, Terminal Bench 2.1: 90.0% vs 84.6%). However, Fable 5 is not part of Fugu's underlying model pool, so this is a comparison against a rival system, not an internal component. All Fugu-Ultra v1.1 benchmarks are vendor-reported and have not been independently reproduced — test on your own workload before switching.
Discussion
0 comments