Why Sakana’s Fugu Shows the Future of AI Is a Manager, Not a Bigger Brain

When the author opened Sakana AI’s newly released Fugu, the first impression was that it was just another large model, yet a closer look revealed it behaves more like a model manager that decides who should handle each sub‑task, how to verify results, and how to stitch the outputs together.

What Fugu Is

Fugu is defined by the official post as a complete multi‑agent orchestration system accessible through a single API. It combines three concepts: a single API, multiple agents, and an orchestration layer. When a request is made, the system may answer directly, split the task, delegate to other models, request verification, and finally aggregate the results.

Technical Foundations

The underlying technology is based on two ICLR 2026 papers, TRINITY and Conductor . TRINITY introduces a lightweight coordinator that assigns roles such as Thinker, Worker, and Verifier to different models across multi‑turn tasks. Conductor enables the system to learn natural‑language collaboration strategies, allowing it to discover when and how agents should communicate and which prompts to use.

How It Works in Practice

From the author’s experience with Claude, Codex, Gemini, and other models, the workflow mirrors manual multi‑model pipelines: one model parses code structure, another checks product logic, a third validates the output, etc. Fugu aims to automate this human‑driven scheduling.

Benchmark Claims

Official benchmark graphics compare Fugu Ultra with Fable 5, Mythos Preview, Opus 4.8, Gemini 3.1 Pro, and GPT 5.5 on three suites:

SWE Bench Pro: Fugu Ultra 73.7, Opus 4.8 69.2, Gemini 3.1 Pro 54.2, GPT 5.5 58.6

TerminalBench 2.1: Fugu Ultra 82.1, Opus 4.8 74.6, Gemini 3.1 Pro 70.3, GPT 5.5 78.2

LiveCodeBench Pro: Fugu Ultra 90.8, Opus 4.8 84.8, Gemini 3.1 Pro 82.9, GPT 5.5 88.4

The author notes that Fable 5 and Mythos Preview are not actually in Fugu’s agent pool, so the comparison mixes different benchmark scopes.

Critiques

Three layers of criticism emerge from the community:

Ability doubts : questions about which benchmarks were used, reproducibility, and whether the results are cherry‑picked.

Product doubts : concerns over pricing, usage limits, latency, regional availability, and refund policies.

Narrative doubts : skepticism about the claim that a scheduling layer solves AI sovereignty and whether it merely replaces one centralised provider with another.

Case Studies

AutoResearch : Fugu Ultra runs a Karpathy‑style AutoResearch loop on an H100 for 14 hours, performing 123 experiments and improving batch size, depth, learning rate, and optimizer, achieving an average BPB of 0.9774.

Financial time‑series : Using an anonymous 50‑week stock series, Fugu Ultra turns $10 000 into $11 943.22 (19.43 % return) over five runs, outperforming three other frontier models that stay below 15 %.

Blind chess : In four consecutive blind‑chess games, Fugu Ultra maintains accurate long‑term state while other models drift, ending each game with a forced win.

Mechanical iris CAD : Fugu Ultra correctly generates a rotating, closing iris mechanism, whereas other models produce gaps or weak connections.

All cases are impressive but not fully publicly reproducible.

Cost, Latency, and Explainability

Pricing tiers are Standard, Pro, and Max at $20, $100, and $200 per month. Ultra’s API costs $5 per million input tokens and $30 per million output tokens, with higher rates beyond 272 K context. The blended‑rate model charges based on the highest‑tier agent used, not per‑agent.

Latency is inherent to multi‑agent orchestration because tasks are split, verified, and aggregated, making it unsuitable for low‑latency chat but acceptable for code review, patent analysis, or security audits.

Explainability is a major concern: if the scheduling layer is closed‑source, users cannot see which agents processed data, how failures were handled, or whether data was sent to disallowed models. The author stresses the need for logs, data‑boundary proofs, and failure‑replay capabilities.

Industry Implications

The article argues that the AI field is shifting from “heroic single‑model” competition to “organizational engineering.” Future competitive advantage may lie in how well a system can manage a pool of models, enforce compliance, and provide transparent orchestration.

Fugu’s narrative ties technical capability to geopolitical and enterprise‑level concerns such as AI sovereignty, vendor lock‑in, and regulatory compliance.

Conclusion

The author believes the direction—moving toward AI systems that act like well‑managed teams—is correct, even though Fugu’s current implementation has notable pitfalls in cost, latency, and openness. The broader question of how to build trustworthy, auditable scheduling layers will become central as more powerful models and agents proliferate.