Sakana Fugu

Sakana Fugu, released June 22, 2026 by Tokyo-based Sakana AI, is a different kind of frontier system: not a single large model but a multi-agent orchestration system shipped as one model. Behind a single OpenAI-compatible API, Fugu is itself a language model trained to delegate — it reads your request and either answers directly or assembles a team of expert models, managing selection, delegation, verification and synthesis internally. The tagline says it plainly: “One Model to Command Them All.”

It ships in two variants. Fugu is the balanced, low-latency default for everyday coding, code review and chatbots; Fugu Ultra maximizes quality on hard, multi-step problems, routing between one and three expert agents depending on difficulty. Sakana says Fugu Ultra stands shoulder-to-shoulder with Anthropic’s restricted Fable 5 and Mythos Preview on the hardest engineering, science and reasoning benchmarks — striking, because neither Anthropic model is even in Fugu’s pool.

The pitch is as strategic as it is technical. A swappable pool is a hedge against single-vendor dependency and export controls — Sakana points directly at the recent restrictions that took Anthropic’s Fable and Mythos models offline. The catch: Fugu is a paid API service with no open weights, orchestration adds tokens to the bill, and it is not available in the EU/EEA at launch.

How Sakana Fugu works

Fugu is not a hand-coded workflow engine — it’s a language model specialized in orchestration. On each request it can solve the problem directly when that is enough, or assemble and coordinate a team of expert models when the task calls for more. Model selection, delegation, agent-to-agent communication, verification and final synthesis all happen internally, so the complexity never reaches your code.

• A swappable agent pool. Fugu routes across a pool of closed and open models, and can call instances of itself recursively. Fugu Ultra coordinates a deeper pool and routes between one and three agents per problem.
• Learned coordination, not fixed rules. Rather than engineers hand-designing which model does what, Fugu learns when to delegate, how agents should talk, and how to combine their work into one reliable answer. This builds on Sakana’s ICLR 2026 research — Trinity and the Conductor (see below).
• Provider resilience. Because the pool is swappable, if one provider restricts access Fugu routes around it. Sakana frames this as a blueprint for “AI sovereignty,” and plans to fold newer models — including open models and its own — into the pool over time.
• Compliance controls. Teams with data, privacy or compliance requirements can opt specific agents out of the pool.

Fugu vs Fugu Ultra

Benchmarks

Sakana’s technical report (Table 1) compares Fugu and Fugu Ultra against the frontier models in its own pool. Baseline scores are the providers’ own reported numbers; the leader in each row is in bold.

On the hardest, leakage-controlled coding split, Fugu Ultra leads every model it routes among:

And on terminal/CLI agents, even the base Fugu — which picks a single model per request — beats the strongest single model in its pool, evidence that the routing decision itself adds performance:

How to read it:

• Wins and losses swap by benchmark — that’s what “shoulder-to-shoulder” means. Fugu Ultra leads on coding, terminal and chart-reasoning tasks, but base Fugu actually wins SciCode, τ³ and long-context, and GPT-5.5 still tops MRCRv2. Against Anthropic’s restricted models (not in the pool), Fable 5 beats Fugu Ultra on SWE-bench Pro (86.0 vs 73.7) and edges it on Humanity’s Last Exam (53.3 vs 50.0).
• It’s system-vs-model. A Fugu score reflects a routed choice or a coordinated team, not one model answering alone — closer to “best achievable with orchestration” than a like-for-like model comparison.
• It’s a moving target. Because Fugu routes among third-party models, results shift as the pool changes; baselines are provider-reported and the SWE task uses the mini-swe-agent harness.

Pricing and cost

Fugu Ultra has fixed, published per-token pricing. The base Fugu model bills differently: when a single agent handles a request you pay that model’s standard rate, and when several agents coordinate Sakana charges a single rate based on the top-tier model involved rather than stacking fees.

Fixed pricing for fugu-ultra-20260615, per 1M tokens.

Cost notes that matter in practice:

• Orchestration tokens count. Fugu’s usage object separates user-visible tokens from orchestration tokens, but orchestration tokens are billed at the same input/output rates. A multi-agent run can consume far more tokens than a single model call — budget for it and cap with max_output_tokens and monitoring.
• Subscriptions vs pay-as-you-go. Plans run $20 (Standard), $100 (Pro), $200 (Max) per month, with usage-based billing for heavier and enterprise workloads. Subscribe before the end of July 2026 and Sakana adds a free second month at your tier.
• Predictability trade-off. You’re paying for orchestration quality, not a flat per-model rate — powerful on hard tasks, but harder to forecast than a single model like Gemini 3.1 Pro.

How to access Sakana Fugu

Fugu is OpenAI-compatible, so you can reuse your existing OpenAI client — just change the base URL and key. Sakana recommends the Responses API; Chat Completions and Models APIs are also supported.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.sakana.ai/v1",
    api_key="YOUR_API_KEY",
)

resp = client.responses.create(
    model="fugu-ultra",
    input="Reproduce the main result of this paper and report what failed.",
    timeout=600.0,   # Ultra runs can be long — raise client-side timeouts
)
print(resp.output_text)

A few practical notes: for the fugu model, set reasoning.effort to high or xhigh/max (any other value is rejected). For complex fugu-ultra tasks, raise client-side timeouts. And note that max_output_tokens on fugu-ultra applies only to the final response — the orchestrator still uses its own limit.

Use cases and worked examples

Fugu’s value shows up less in single-prompt answers and more in long, messy, multi-step workflows — exactly what Sakana focused on with its ~500-user beta.

1. Automated, long-horizon research

Beta users ran Fugu in a near-fully-automated research mode (Sakana calls it AutoResearch): it explores ideas, runs experiments, interprets failures, revises its approach, and keeps making progress over time with little human intervention. This open-ended, multi-step work is the canonical Fugu Ultra task — and where Sakana reports it beating Gemini 3.1 Pro, Opus 4.8 and GPT-5.5.

2. End-to-end code review

“For code review, Fugu Ultra is significantly better than GPT-5.5… Where other tools flag about three issues, Fugu surfaced more than twenty. It’s become the model I run all my reviews through.” — Software Engineer (Sakana beta)

Because review fans out across multiple expert agents that cross-check each other, Fugu tends to surface bugs a single pass misses.

3. One-shot security assessments

“Given one scoped instruction, Fugu drove a full security assessment end-to-end — recon, XSS/SQLi checks, auth review, and a clean report with evidence and retest steps — staying inside scope and avoiding destructive actions.” — Cyber Security Engineer (Sakana beta)

Reviewers also noted unusually strong persona stability across long sessions — useful for agent products where other models tend to drift.

How Sakana Fugu compares

• vs the models it routes to (GPT-5.5, Gemini 3.1 Pro, Claude Opus 4.8): Fugu doesn’t replace them — it conducts them. Sakana reports the orchestrated system beating each of these base models (at their high/max/xhigh settings) on applied tasks.
• vs Anthropic’s restricted frontier (Fable 5, Mythos Preview): the headline claim is parity on the hardest benchmarks without those models in the pool — and without the export-control exposure that took Mythos offline days after launch.
• vs traditional multi-agent frameworks (LangGraph, AutoGen, CrewAI): those make you design and maintain the agent graph; Fugu ships orchestration as a learned model behind one API, so there’s no glue code or per-task workflow engineering.

Short version: reach for Fugu when a task is long, messy and multi-step, or when vendor resilience matters; call a single model directly when latency and cost predictability matter more. For the wider field, see our 2026 LLM landscape guide.

Limitations and known gotchas

• No EU/EEA access at launch. The API is restricted there pending GDPR compliance.
• Not open-weight. Fugu is a hosted API that depends on the third-party providers it routes to; you can’t self-host it.
• Cost is harder to predict. Orchestration tokens and multi-agent runs can balloon usage; cap with max_output_tokens, lean on cached input, and monitor the usage breakdown.
• Ultra latency. Hard tasks can run long enough to need raised client timeouts — it’s built for depth, not snappy interactive turns.
• Vendor-reported, system-vs-model benchmarks. The parity claims exclude the very Anthropic models they’re measured against; verify against your own workload.
• Young product. GA on June 22, 2026 after a ~500-user beta (the small variant was earlier called “Fugu Mini”) — expect rough edges and an evolving pool.

The research behind Fugu

Fugu is grounded in two papers accepted at ICLR 2026, plus a technical report:

• TRINITY: An Evolved LLM Coordinator — a lightweight, evolved coordinator that assigns models to Thinker / Worker / Verifier roles across multiple turns, adapting dynamically to the task.
• Learning to Orchestrate Agents in Natural Language with the Conductor — uses reinforcement learning to discover natural-language coordination strategies, learning how to prompt and route agents instead of relying on hand-designed workflows.
• Sakana Fugu Technical Report (2026) — the system details and benchmark methodology.

Sakana’s orchestration pedigree predates Fugu: its ALE-Agent placed 21st out of 1,000 human experts in a coding competition.

Availability and roadmap

Sakana Fugu is generally available today (June 22, 2026) through console.sakana.ai, following a beta with roughly 500 users from late April 2026. The fugu-ultra-20260615 ID pins a specific Fugu Ultra version, while fugu and fugu-ultra track the latest.

Sakana frames the launch as “a starting point, not a finish line.” Because Fugu is built on learned orchestration rather than fixed workflows, it improves as the ecosystem does: the roadmap is to expand the expert-agent pool (including open models and Sakana’s own), strengthen coordination for long-running agentic tasks, and give users more control over how Fugu routes on their behalf.

Best for

• Long-horizon automated research and paper reproduction
• End-to-end code review and security assessments
• Vendor-resilient deployments that must route around access disruptions

Pros & cons

Frequently asked questions

Sources

• Sakana AI — Sakana Fugu: One Model to Command Them All
• Sakana AI — Fugu product page
• Sakana AI — Console: Models
• Sakana AI — Console: Pricing
• The Decoder — Sakana AI's Fugu orchestrates multiple LLMs (benchmark Table 1)
• ITmedia — Sakana AI、一部「ミュトス越えの性能」うたうAIを提供
• sakutto — What Is Sakana Fugu? Performance vs Fable 5 (Fable comparison figures)