Still debating whether to use GPT, Claude, or Gemini?
On June 22, 2026, Japanese AI unicorn Sakana AI dropped a bombshell. Instead of launching yet another bigger monolithic model, they released something fundamentally different — Sakana Fugu, an orchestrator model trained to command other models.
This strategy directly challenges the industry’s established rules. While everyone else is racing to stack parameters and hoard GPUs, Sakana AI chose a radically different path: instead of dominating with a single model, let multiple top-tier AIs fight together as a team.
What Is Fugu? Not a Model, But a Conductor
Put simply, Sakana Fugu is itself a language model, but it does not answer your questions directly. Its job is to decide who answers.
When you send a request, Fugu automatically analyzes the task nature, then picks the best model combination from its agent pool. Complex coding problem? Call the software engineering specialist. High-difficulty math reasoning? Switch to the math-strong model. Need multi-step verification? It auto-assembles a squad with Thinkers, Workers, and Verifiers.
The entire process is transparent to you. You only need to call a single OpenAI-compatible API endpoint. Model switching, task decomposition, cross-validation — all handled by Fugu automatically.
Sakana AI positions this as the realization of collective intelligence. They believe the next frontier of AI is not infinite scaling of a single model, but identifying, combining, and amplifying the complementary strengths of many models.
Two Versions, Two Battlefields
Fugu currently offers two variants, designed for different scenarios:
Fugu (Standard): Balances performance and latency for everyday coding, code review, and chatbot interactions. You can opt specific agents out of the pool to meet data privacy or compliance requirements.
Fugu Ultra (Flagship): Coordinates a deeper pool of expert agents, designed for the hardest, highest-stakes problems. Early users already deploy it for Kaggle competitions, paper reproduction, security analysis, and patent investigation.
Both variants share the same API endpoint. Switch between them based on workload without changing your integration code.
Hard Data: Is It Really Better Than Opus 4.8?
According to official benchmark data published by Sakana AI, Fugu Ultra delivers impressive results across multiple authoritative tests:
| Benchmark | Fugu Ultra | Opus 4.8 | Gemini 3.1 Pro | GPT 5.5 |
|---|---|---|---|---|
| SWE-Bench Pro | 73.7 | 69.2 | 54.2 | 58.6 |
| LiveCodeBench | 93.2 | 87.8 | 88.5 | 85.3 |
| GPQA-Diamond | 95.5 | 92.0 | 94.3 | 93.6 |
| Humanity’s Last Exam | 50.0 | 49.8 | 44.4 | 41.4 |
| TerminalBench 2.1 | 82.1 | 74.6 | 70.3 | 78.2 |
SWE-Bench Pro is the gold standard for measuring software engineering capability. Fugu Ultra’s 73.7 beats Opus 4.8’s 69.2. GPQA-Diamond (graduate-level science Q&A) hits 95.5, the highest score among publicly accessible models.
Sakana AI claims Fugu Ultra stands shoulder-to-shoulder with Anthropic’s Fable 5 and Mythos Preview on engineering, science, and reasoning benchmarks. Notably, Fable 5 and Mythos Preview are not in Fugu’s agent pool because they are not publicly accessible.
The Tech Behind: Evolutionary Algorithms + Reinforcement Learning
Fugu’s core technology comes from two ICLR 2026 papers: TRINITY and Conductor.
TRINITY proposes a lightweight evolved coordinator. Using a head with only about 10K parameters, it dynamically assigns Thinker, Worker, and Verifier roles to selected LLMs. The coordinator itself is optimized through evolutionary algorithms, without relying on hand-crafted rules.
Conductor trains a conductor through reinforcement learning, letting it discover optimal agent communication patterns and collaboration strategies. Research found that RL-trained coordination strategies are often counter-intuitive yet highly efficient — patterns human designers would unlikely conceive.
This means Fugu’s scheduling logic is not a hardcoded pipeline, but learned. It dynamically assembles teams from the model pool and coordinates them through non-obvious but highly efficient collaboration patterns.
More uniquely, Fugu can recursively call itself. If the first answer is insufficient, it re-evaluates coordination strategy and relaunches the workflow. This is positioned as a new form of test-time scaling — adjusting compute at inference time without retraining.
Real-World Cases: From Ancient Texts to Stock Trading
Sakana AI showcased six concrete cases demonstrating Fugu’s real-world capabilities:
AutoResearch: Fugu Ultra autonomously ran 123 experiments over ~14 hours on a single H100 GPU, achieving the best BPB score (0.9774), surpassing all frontier competitors.
Japanese Historical Text Recognition: When restoring reading order of classical Japanese kana letters, Fugu achieved NED 0.80, while competing models scored only 0.24 or failed entirely.
Rubik’s Cube Solver: Fugu-generated pure Python code solved all 300 test cubes, never losing in move count (7 wins, 293 ties, 0 losses).
CAD Mechanical Design: Designed a working mechanical iris, while other models produced designs with gaps or incomplete structures.
Blindfold Chess: Maintained perfect accuracy across four consecutive games.
Stock Trading: Achieved +19.43% average return over a 50-week backtest, while other frontier models remained below 15%.
Pricing and Business Model
Fugu offers both subscription and pay-as-you-go pricing:
Subscription (includes both Fugu and Fugu Ultra):
- Standard: $20/month
- Pro: $100/month (10x Standard quota)
- Max: $200/month (20x Standard quota)
Subscribe before July 2026 and get the second month free.
Pay-as-you-go (Fugu Ultra):
- Input: 10 beyond 272K context)
- Output: 45 beyond 272K context)
- Cached input: $0.50/1M tokens
This pricing is close to Opus 4.8 and significantly cheaper than Fable 5.
The Biggest Selling Point: Avoiding Single-Vendor Lock-in
Fugu’s narrative this time is not just about performance, but AI sovereignty.
The underlying model pool can be swapped at any time. If a supplier gets cut off by export controls, raises prices, or throttles access, Fugu dynamically switches to alternative models without service interruption. Sakana AI positions this as a new-era AI sovereignty solution.
For organizations deploying AI in critical infrastructure, finance, or government, this is extremely compelling. You no longer need to worry about Anthropic suddenly restricting overseas access or an API getting shut down without warning.
Controversies and Reality
Of course, Fugu is not without criticism.
Some developers point out that Fugu itself is a closed-source scheduler relying on closed-source foundation models. You used to at least know which model you were using; now you have no idea which model was called, how many tokens were consumed, or what the cost was — completely black-box. Some bluntly call it just an AI service wrapper.
Real-world tests also show that while Fugu Ultra excels at complex tasks, its cost-effectiveness on simpler tasks may lag behind using a single model directly. One team found Fugu was roughly 17x more expensive than GLM 5.2, though output quality was higher.
Additionally, Fugu’s theoretical capability ceiling depends on the strongest model in its agent pool. Without top-tier models available, even the most brilliant orchestration cannot break through the ceiling.
Conclusion: 2026, the Year of Orchestrator Models?
The emergence of Sakana Fugu marks the AI competition officially entering phase two.
For the past three years, the keywords of the global AI arms race were parameters, compute, and single-model benchmarks. Now, Fugu proves the viability of an alternative path: instead of pursuing an invincible single model, orchestrate and integrate diverse AIs globally, leveraging each’s strengths while mitigating weaknesses.
The advantages are clear: flexible, controllable, lower risk, immune to vendor throttling or regulation. For developers and enterprises, this means no longer betting on a single vendor, but gaining collective intelligence加成 through a unified interface.
Whether Fugu ultimately disrupts the existing landscape or not, it has at least proven one thing: the future of AI belongs not only to the biggest model, but also to the smartest orchestrator.
If you are interested in multi-agent systems or AI orchestration technology, feel free to discuss in the comments. Do you think conductor models will become the next mainstream trend?
Comments