Sakana Fugu AI - Japanese Mythos5/Fable5 by Sakana AI

Sakana Fugu AI is an orchestration system from Tokyo-based AI lab Sakana AI — not a traditional single model, but a "model of models" that breaks your request into pieces, hands each piece to whichever frontier AI in its pool fits best, and fuses the results into one answer. Launched June 22, 2026, it represents a genuine shift in how AI products are built: instead of training a bigger model, Sakana trained a conductor. This is an independent guide — we're not affiliated with Sakana AI.

What Fugu actually is

Every request hits Fugu's conductor first. It decides whether to answer directly or assemble a multi-agent team — Thinker, Worker, Verifier roles — drawn from a pool of frontier models that has included GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro. The Verifier step is what separates Fugu from a simple router: before returning an answer, an agent checks the output for errors and sends it back for revision if needed. You interact with one OpenAI-compatible API; the orchestration happens invisibly underneath. Read our comparison of Fugu vs monolithic models for why 2026 became the year of orchestrated intelligence.

Quick facts

• Product: Sakana Fugu — multi-model orchestration via a single API
• Maker: Sakana AI (Tokyo-based AI lab)
• Launched: June 22, 2026
• Tiers: Fugu (fast, everyday work) and Fugu Ultra (quality-first, hard multi-step tasks)
• Later releases: Fugu Ultra v2 and Fugu Max arrived September 11, 2026
• API: OpenAI-compatible endpoint; 1M token context window
• Availability note: not available in the EU/EEA at launch (GDPR compliance pending)

The two tiers (now three)

At launch, Fugu was the lower-latency default for everyday work and Fugu Ultra the quality-first variant for harder multi-step tasks where answer depth matters more than response time. In September 2026, Sakana added Fugu Max — a cheaper, faster sibling built more heavily on open-weight and specialist models, priced at $2 input / $6 output per million tokens. Fugu Ultra's pay-as-you-go pricing sits at $5 input / $30 output per million tokens, with subscriptions from $20 to $200 a month. The tier you choose is really a choice about your workload: straightforward tasks don't benefit from orchestration, while multi-step agentic work is where the conductor earns its keep.

Why orchestration instead of a bigger model

For years, AI competition centered on building bigger models. Fugu's premise is different: the consequential capability may lie in teaching AI to direct other AI. The design draws on two Sakana papers accepted at ICLR 2026 (on learned orchestration), and the commercial argument is pragmatic — a swappable pool means no dependency on any single provider. Sakana has been explicit that single-vendor dependency is "a material vulnerability," a point sharpened by export-control restrictions affecting some frontier models. If one model in the pool becomes unavailable, the conductor reroutes. A monolithic model has no such fallback.

Benchmarks (with the usual caveats)

Sakana reports Fugu Ultra scoring 73.7 on SWE-Bench Pro — ahead of GPT-5.5 (58.6) and Claude Opus 4.8 (69.2) as standalone models — plus strong numbers on TerminalBench, LiveCodeBench, and GPQA-Diamond. Independent testers have run their own comparisons with mixed-but-generally-positive results, particularly on debugging and multi-step coding tasks where the Verifier role catches errors single models miss. The honest caveat: these are Sakana's internal numbers and haven't been independently verified at the benchmark level. Enterprises should validate cost, performance, and reliability against their own workloads before committing.

Who it's for

Developers and teams doing multi-step agentic work: software engineering tasks, complex research pipelines, workflows where combining approaches measurably helps. For straightforward chat or summarization, a single cheaper model does the same job for less — orchestration overhead isn't free. See our guide on integrating Fugu into a development workflow. And for the backstory, how Sakana AI built Fugu covers the lab's path from research to product.

The export-control angle

Fugu's launch timing wasn't accidental. It arrived as export-control restrictions were reshaping who could access which frontier models — and Sakana positioned orchestration as the answer to single-vendor dependency being "a material vulnerability." The pitch: if your AI stack depends on one provider's model and that model becomes restricted, repriced, or deprecated, you're exposed. A conductor with a swappable pool routes around the problem. Whether you buy the geopolitics or not, the engineering point stands — abstraction over providers is more resilient than dependence on one.

This is also why Fugu's pool composition matters less than its architecture. The specific models in the pool will change — they already have, with Fugu Max leaning on open-weight and specialist models like NVIDIA's Nemotron family. The durable asset is the conductor: the learned orchestration that decides which model handles which piece. Models commoditize; coordination compounds.

Fugu Max and the pricing war

The September 2026 launch of Fugu Max (with Fugu Ultra v2) moved the battlefield from model inference costs to orchestration economics. At $2 input / $6 output per million tokens, Fugu Max undercuts the output costs of several frontier models by 40–60% — while reportedly beating Fugu Ultra v2 on six of Sakana's own benchmarks for tool-use and terminal-style tasks. The message to the market: you don't need the most expensive model for every subtask; you need a conductor smart enough to route easy subtasks to cheap models and hard ones to the best available. That's the arbitrage Fugu is selling.

Frequently asked questions

Is Fugu a single AI model? No — it's an orchestration system that coordinates a pool of frontier models behind one API. You get one response; the conductor decides which models contributed.

What's the difference between Fugu and Fugu Ultra? Fugu is the faster, lower-latency tier for everyday work; Fugu Ultra is the quality-first tier for hard, multi-step tasks. Fugu Max (Sept 2026) is the cheaper, faster sibling.

Can I use it in the EU? Not at launch — Sakana held back EU/EEA availability pending GDPR compliance work.

Do I need to manage the individual models? No. You call one API and pick the tier; the conductor handles model selection internally.

Is this affiliated with Sakana AI? No — independent guide. For the product itself, use Sakana AI's official channels.

The bottom line: Fugu bets that the future of AI isn't one giant model but a conductor smart enough to direct many. Whether orchestration becomes the decisive frontier is still an open question — but as a commercial product, it's already forcing everyone to think about what "a model" even means.