Sakana Fugu Review: When Multi-Agent Orchestration Actually Pays Off

Sakana Fugu is a model that decides for you whether to answer a prompt directly or convene a team of frontier models. The launch numbers are dazzling and every one of them is vendor-reported. Here is what Fugu actually is, why its headline claims are hard to verify, and a first-party benchmark that answers the question every other review dodges: when is the orchestration overhead worth paying for?

What Sakana Fugu is

Sakana AI, the Tokyo lab, launched Sakana Fugu on June 22, 2026. It is a multi-agent orchestration model that behaves like a single model: you send a request to one OpenAI-compatible endpoint and Fugu decides whether to answer directly or assemble and coordinate a team of expert frontier models, then verify and synthesize the answer. At launch it is live globally except the EU/EEA, and is also reachable through OpenRouter and Vercel.

The key detail people miss is that Fugu is itself a language model trained to call a pool of other LLMs, and instances of itself. It routes, delegates, verifies, and synthesizes as a learned behavior, not through hand-coded if/else rules. There are two variants. Fugu is the balanced tier with lower latency and supports selective opt-out of specific agents or providers for compliance. Fugu-Ultra targets maximum quality on hard problems, runs a fixed pool with no opt-out, and carries the model id fugu-ultra-20260615.

How the orchestration works

Fugu builds on two ICLR 2026 papers on learned orchestration: Trinity, a lightweight evolved coordinator that assigns Thinker, Worker, and Verifier roles, and The Conductor, reinforcement-trained to discover natural-language coordination strategies. The technical report is arxiv.org/abs/2606.21228, with authors led by Yujin Tang and Edoardo Cetin. In practice, Fugu can call frontier models such as GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.8, coordinate their outputs, and return a single synthesized response.

This is the appeal and the catch at once. The routing is proprietary and hidden by design. You cannot see which model actually answered your query and you cannot steer it. On the Fugu tier you can opt whole providers or model categories out, but you cannot pick the model that serves a given request. If you want to understand where Fugu sits relative to the app-level pattern developers already use, our write-up of the orchestrator-executor pattern is the right companion.

The vendor-reported numbers

Every figure in this section is Sakana-reported and, as of July 2026, has not been independently reproduced. Read them as vendor claims, not established fact. Sakana states that Fugu and Fugu-Ultra top 10 of 11 major benchmarks and outperform the very frontier models they orchestrate. Sample claimed numbers:

• SWE-Bench Pro: Fugu-Ultra 73.7 vs Claude Opus 4.8 at 69.2 (Sakana-reported)
• GPQA-Diamond: 95.5 · LiveCodeBench: 93.2 (Sakana-reported)
• Humanity's Last Exam: 50.0 · MRCRv2: 93.6 (Sakana-reported)

Sakana also claims Fugu-Ultra is shoulder-to-shoulder with Anthropic's Fable 5 and Mythos Preview. Those two models are explicitly not in Fugu's agent pool: on June 12, 2026 the US government, citing national-security authorities, imposed export controls on them and Anthropic suspended access, making them inaccessible to foreign nationals worldwide (the controls were later lifted in early July). Sakana CEO David Ha frames the swappable pool as a hedge, arguing that relying on one company's APIs for critical infrastructure is a material vulnerability and that the pool routes around provider disruptions without code changes.

Cost and latency, by design

Fanning one request across several models plus a verification pass is not free, and the cost shows up in both dollars and seconds. These specifics come from independent hands-on reviewers (TechTimes, ayautomate, buildfastwithai), not Sakana documentation, so treat them as indicative:

• Up to roughly $10 per message on heavy tasks.
• A light question clocked near 108 seconds, with about 60% of billed tokens spent on hidden back-channel orchestration.
• Coding tasks stretching toward 30 minutes.

That trade-off is intrinsic to the design, not a bug. The question is whether the extra quality you buy justifies it, and that depends entirely on the task. If you are assembling reliability at the infrastructure layer instead of the model layer, our LLM routing and failover guide covers the version you control yourself.

Why the claims are hard to verify

Fugu has a structural verifiability problem, and it is worth naming plainly. First, the benchmarks are Sakana-reported and not independently reproduced. Second, the orchestrator and its model mix are closed, with an undisclosed open-versus-closed split, so no one outside Sakana knows what actually ran. Third, the named parity baselines, Fable 5 and Mythos Preview, were pulled from access, which means no head-to-head was possible; Sakana compared against provider-reported reference scores. Community sentiment, approximated through secondary coverage rather than login-gated social feeds, skews skeptical, with the recurring question being whether Fugu is just a router or wrapper.

None of that proves Fugu is weak. It means the honest posture is agnostic: the headline claims are plausible and unverified, and a closed orchestrator that hides its routing is difficult to audit even in principle.

Fugu vs a gateway vs orchestrator-executor

Fugu is often lumped in with two things it is not. A plain LLM gateway and the orchestrator-executor pattern solve overlapping problems in different layers. The honest differences:

So a gateway routes to models you pick and tells you who served the request; the orchestrator-executor pattern makes you hardcode coordination in your own app; Fugu internalizes coordination into the model behind one endpoint and takes the steering wheel. Fugu's convenience is real, and so is the loss of control. These are complementary categories, not equivalents.

When orchestration actually pays

Every review gestures at Fugu being slow and pricey, but none answers the real question: when is the overhead worth it? Sakana frames Fugu as universally superior. That framing collapses the moment you look at first-party executed data instead of vendor benchmarks.

In July 2026, DataLLM Lab ran an executed coding benchmark: 13 models, 9 tasks that generate code and then run it against hidden tests. 10 of 13 models scored a perfect 9/9. Quality converged on standard tasks, across an 88x cost spread from $0.10 to $8.83 per 1,000 tasks.

Layer that onto Fugu's economics. If a single cheap model already solves the standard task perfectly, then Fugu-Ultra's multi-model fan-out plus verification pass, up to ~$10 per message with ~60% back-channel tokens and up to 30-minute latency, is pure overhead with no quality to capture. Orchestration only pays where quality has not converged: the frontier of genuinely hard, non-standard problems, which is exactly where Sakana's own showcase examples live (Kaggle competitions, paper reproduction, cybersecurity, patent search).

So the decision rule is simple and, notably, absent from every competitor article because none has executed-benchmark data showing the convergence: match task difficulty to routing complexity. Use a cheap single model for the converged 80% of standard work, and reserve an orchestrator for the long-tail 20% of truly hard problems. If you are choosing models for agents, our roundup of the best LLMs for AI agents in 2026 and the broader best LLM 2026 guide apply the same rule.

FAQ