The End of Monolithic AI? Sakana Fugu Unleashes Multi-Agent Orchestration
Answer-first verdict: Sakana AI's Fugu and Fugu Ultra are not traditional large language models (LLMs) but rather multi-agent orchestration systems designed to dynamically coordinate a diverse pool of AI models. This innovative approach allows them to achieve and, in some cases, surpass the benchmark performance of leading frontier LLMs like Anthropic's Fable 5, Mythos Preview, OpenAI's GPT-5.5, and Google's Gemini 3.1 Pro by intelligently delegating tasks to specialized agents.
TL;DR / at-a-glance:
- Sakana AI's Fugu is a multi-agent orchestration system, not a single LLM.
- It dynamically calls and coordinates specialized AI models for complex tasks.
- Fugu Ultra matches or exceeds top-tier models on benchmarks like SWE-Bench Pro and LiveCodeBench.
- This approach offers frontier performance without direct access to export-controlled models.
- Real-world use cases show strong performance in complex, multi-step workflows.
- Pay-as-you-go on Fugu Ultra is $5 per 1M input tokens and $30 per 1M output tokens; subscriptions start at $20/month.
- Not available in the EU or EEA. A complex Ultra request can take 15–25 minutes.
- All benchmark figures are vendor-reported and not independently reproduced.
- Last verified: 2026-08-01
What is Sakana Fugu and How Does it Work?
Sakana Fugu is a paradigm shift in how we approach frontier AI. Instead of relying on a single, massive LLM to handle all tasks, Fugu acts as an intelligent coordinator. It's a small language model itself, specifically trained to understand a request, then decide which other specialized AI models (agents) to call upon, how to route subtasks, and finally, how to synthesize their outputs into a coherent, high-quality answer. This is an application of a broader industry shift covered in our deep dive into multi-agent orchestration systems. This entire process happens internally and is invisible to the end-user, who interacts with Fugu as if it were a single, powerful model via an OpenAI-compatible API.
This dynamic orchestration allows Fugu to leverage the strengths of various models, avoiding the limitations inherent in monolithic architectures. It can even recursively call instances of itself, enabling deeply nested reasoning chains for exceptionally complex problems.
The important detail is that the orchestrator is a trained model, not a prompt-engineering script. The architecture comes out of two ICLR 2026 papers from Sakana AI: TRINITY, an evolved LLM coordinator, and Conductor, on learning to orchestrate agents in natural language. The orchestrator was trained with reinforcement learning to discover natural-language coordination strategies — effectively learning how agents should talk to each other, and what focused prompts make a diverse pool of LLMs outperform any single worker. That is what separates Fugu from a model router like OpenRouter, which forwards a request to a model you picked: Fugu adapts its strategy per task.
Fugu vs. Fugu Ultra: Tailored for Performance and Complexity
Sakana AI offers two variants of its orchestration system:
- Fugu: Tuned for a balance of strong performance and low latency, Fugu is ideal for everyday workflows such as coding assistance, code review, and interactive chatbots. It also allows for configurable agent exclusion lists, catering to specific privacy or compliance requirements.
- Fugu Ultra: Optimized for maximum answer quality on difficult, multi-step problems. Fugu Ultra employs a deeper agent pool and has been deployed successfully in demanding tasks like AI research, scientific paper reproduction, cybersecurity analysis, and large-scale patent investigations.
| Feature | Fugu | Fugu Ultra |
|---|---|---|
| Optimised for | Latency, everyday work | Quality, complex multi-step tasks |
| Agent pool | Configurable (opt-out supported) | Fixed (full pool) |
| Benchmark leader | 92.9 LiveCodeBench | 93.2 LiveCodeBench, 73.7 SWE-Bench Pro |
| Pricing model | Dynamic (rate of the top model used) | Fixed $5/$30 per 1M tokens |
| Best for | Coding, chat, code review | Research, security, long-horizon tasks |
Fugu Ultra's pool is fixed precisely because it relies on the full agent set to deliver its benchmark performance; the opt-out list only exists on the standard tier.
How Does Fugu Ultra Match or Beat Frontier LLMs on Benchmarks?
Sakana AI's Fugu Ultra has demonstrated impressive benchmark results, often matching or outperforming established frontier models across various domains. It represents a different approach to the ones used by tools in our AI Coding Assistant Showdown.
- SWE-Bench Pro: Fugu Ultra scored 73.7, surpassing Opus 4.8 (69.2) and GPT-5.5 (58.6) in this demanding software engineering benchmark.
- LiveCodeBench: Fugu Ultra achieved 93.2, outperforming Gemini 3.1 Pro (88.5) and Claude Fable 5 (89.8).
- Humanity's Last Exam: Fugu Ultra reached 50.0, closely matching Opus 4.8's 49.8.
- GPQA-D: Fugu Ultra scored 95.5, equaling regular Fugu and notably exceeding Mythos Preview's 94.6.
The full scoreboard from Sakana's published data, including where Fugu loses:
| Benchmark | Fugu Ultra | Fugu | Best competitor |
|---|---|---|---|
| SWE-Bench Pro (agentic coding) | 73.7 | — | Fable 5: 80.0 (restricted) · Opus 4.8: 69.2 · GPT-5.5: 58.6 |
| LiveCodeBench v6 (coding) | 93.2 | 92.9 | Fable 5: 89.8 · Gemini 3.1 Pro: 88.5 |
| GPQA-Diamond (graduate science) | 95.5 | 95.5 | Mythos Preview: 94.6 |
| Humanity's Last Exam | 50.0 | — | Fable 5: 53.3 · Opus 4.8: 49.8 |
| MRCRv2 (long-context recall) | 93.6 | — | GPT-5.5: 94.8 |
| CTI-REALM (cybersecurity) | 69.4 | — | Opus 4.8: 69.6 |
These are vendor-reported figures from Sakana AI and have not been independently reproduced by third-party labs as of August 2026. Treat them as a claim to test on your own workloads.
The pattern is clear. Fugu Ultra wins on decomposable, multi-step, agentic tasks where combined model strengths matter, and trails on tightly-constrained single-domain brute-force reasoning — long-context recall, cybersecurity, the hardest agentic coding — where the largest standalone models still hold an edge provided you can access them. The benchmark Fugu most conspicuously loses, SWE-Bench Pro to Fable 5 at 80.0, is one where the winning model was pulled from public access on June 12, 2026 under U.S. export controls.
That is the strategic point: Fugu reaches frontier-level performance without depending on models subject to export controls, which matters for anyone thinking about AI sovereignty.
Why does multi-agent orchestration beat a single model?
Every AI model you use has the same hidden limit: one brain does the whole job. The same neural network plans your architecture, writes your code, debugs your errors, and designs your UI. When the job gets big, that one brain starts dropping things — context gets lost, the graphics land but the controls break, the physics works but part of the application never renders. We have all had a long build come back 90% right and 10% broken.
Multi-agent orchestration attacks that differently. An orchestrator decomposes the task, routes each sub-task to a specialised expert, and fuses the results. If one expert makes a mistake, another covers for it. It mirrors how human teams work: a senior architect breaks down a project, assigns pieces to specialists, reviews the work, synthesises the final product.
For builders working with AI agent operating systems, Fugu's architecture is a real alternative to hand-wiring multi-model pipelines. Instead of building custom orchestration with something like Macaron-V1's mixture-of-LoRA approach, Fugu packages the orchestration as the model. Teams focused on serving AI agents at scale will recognise the pattern: offloading coordination to a trained model simplifies the deployment surface.
How fast is Fugu Ultra?
It is not fast, and that is deliberate. A complex task that fans out to multiple expert models and runs verification loops can take 15–25 minutes of orchestrated thinking. The trade is answer quality over latency on hard problems where a single model call produces 90% correct output with 10% broken. For everyday coding or chat, the standard Fugu tier is the better choice. The practical pattern is to run Ultra in the background on one-off, high-stakes tasks while you keep working on a faster model.
What does Sakana Fugu cost?
Subscriptions start at $20/month (Standard), $100/month (Pro, 10x usage) and $200/month (Max, 20x usage for continuous tasks). On pay-as-you-go, Fugu Ultra is $5 per million input tokens and $30 per million output tokens, with $0.50 cached input, and rates rising to $10/$45 per million once context exceeds 272K tokens. Standard Fugu uses dynamic pricing: it charges the single rate of the highest-tier underlying model involved and does not stack multi-agent fees on top.
| Pricing tier | Input (per 1M) | Output (per 1M) | Cached input |
|---|---|---|---|
| Fugu Ultra standard | $5.00 | $30.00 | $0.50 |
| Fugu Ultra, context > 272K | $10.00 | $45.00 | $1.00 |
| Fugu Cyber | $6.00 | $36.00 | $0.60 |
The caveat that catches people: Fugu Ultra bills orchestration tokens. The background tokens consumed when Fugu delegates sub-tasks, verifies code and routes between agents count toward your request at standard rates — the provider does not absorb them. A task with heavy fan-out and verification can cost far more than its visible output length suggests.
Against other frontier models at headline rates:
| Model | Input/1M | Output/1M | Combined |
|---|---|---|---|
| DeepSeek-V4-Flash | $0.14 | $0.28 | $0.42 |
| Claude Opus 4.8 | $5.00 | $25.00 | $30.00 |
| GPT-5.5 | $5.00 | $30.00 | $35.00 |
| Sakana Fugu Ultra | $5.00 | $30.00 | $35.00 |
| Claude Fable 5 / Mythos 5 | $10.00 | $50.00 | $60.00 |
Fugu Ultra matches GPT-5.5 and costs half of the restricted Fable 5 — but the effective rate depends on how many orchestration tokens a given task burns.
How do you get access?
Two paths:
- Direct from Sakana AI. Create an API key at console.sakana.ai. The API is OpenAI-compatible at
https://api.sakana.ai/v1, so existing OpenAI SDK code works with a base-URL change. - Through OpenRouter. Fugu Ultra is listed on OpenRouter at the same $5/$30 rate, which is the recommended route where direct access is blocked.
Is Sakana Fugu available in the EU?
No. Sakana explicitly does not provide Fugu services to users in EU or EEA member states while it works to align its black-box routing architecture with GDPR requirements. UK users have also reported access issues. Elsewhere, access may be limited by network conditions or local regulation. Where you are blocked, the OpenRouter path may work.
When should you use Fugu Ultra instead of a single model?
Good fits:
- Long-horizon research — paper reproduction, patent searches, multi-source analysis where sustained progress across many steps matters
- Difficult code review — complex refactoring, security analysis, architecture audit where multiple expert perspectives catch more issues
- Multi-step engineering — where one model returns 90% correct output with 10% broken across different dimensions
- High-stakes one-shot tasks — where being right first time is worth 25 minutes
Bad fits:
- Simple coding or chat (use standard Fugu)
- Budget-sensitive workloads (DeepSeek-V4-Flash at $0.42/M combined serves most routine tasks)
- Tight latency requirements (fan-out adds minutes, not seconds)
- EU/EEA-based teams (access blocked)
What are the alternatives?
| Alternative | How it works | Cost | Best for |
|---|---|---|---|
| OpenRouter Fusion | Fans one prompt out to N models, uses a judge model to synthesise | Per model plus judge | Budget-conscious multi-model fusion |
| Fusion Engine (open source) | Self-hosted Fusion alternative with custom judge prompts | Per-model API costs only | Full control over panel and judge |
| DIY orchestration (LangGraph / AutoGen) | Build your own routing, delegation and synthesis | Engineering time plus API costs | Specific workflow needs |
| Standard Fugu | Same orchestrator, lighter agent pool | Dynamic (top model rate) | Everyday coding and chat |
DIY is more viable now that the orchestration research is public as ICLR 2026 papers. For teams already managing AI token cost optimisation, building your own layer may be more cost-effective at scale.
What are the Practical Implications of Multi-Agent Orchestration?
The real-world applications of Fugu's multi-agent orchestration extend beyond benchmarks. Early beta users have reported:
- Enhanced Problem Solving: In production codebase analysis, Fugu surfaced over 20 issues where competing single models identified only 3, demonstrating its ability to catch diverse bugs simultaneously.
- Consistent Persona Stability: For agent products, Fugu exhibited unusually strong persona stability across long sessions, maintaining its identity where other models tend to drift.
- End-to-End Task Automation: In cybersecurity, Fugu successfully drove full security assessments from a single instruction, including reconnaissance, vulnerability checks, and report generation, all while staying within scope.
These examples highlight the shift from simply building larger models to developing smarter coordination mechanisms, enabling more robust, reliable, and versatile AI systems.
What this means for you
For developers and businesses, Sakana Fugu represents a new avenue for accessing cutting-edge AI capabilities. It suggests that future advancements may not solely come from ever-larger foundational models, but from sophisticated orchestration layers that can dynamically assemble and manage specialized agents. As models solve harder problems, the bottleneck shifts to human task imagination—the ability to frame the right multi-step projects for these orchestrators to tackle. This approach could lead to more efficient, cost-effective, and adaptable AI solutions, particularly for complex, multi-step problems that benefit from diverse expertise.
The people winning with AI in 2026 are not picking one model and sticking with it. They run a stack and route each job to whichever brain — or team of brains — fits it. For one-off, high-stakes work where quality outweighs speed and cost, Fugu Ultra is the strongest publicly accessible option; for everyday work, standard Fugu or a cheaper single model is fine. Own the whole aquarium, not just your favourite fish.
For teams already running automated AI research workflows or building revenue loops with AI agents, Ultra fits as the high-quality layer for complex reasoning.
Related reading
FAQ
Q: Is Sakana Fugu a traditional large language model? A: No, Sakana Fugu is not a single large language model in the traditional sense. It is a multi-agent orchestration system that uses a smaller LLM to coordinate and call upon other specialized AI models to complete tasks.
Q: How does Fugu achieve its high benchmark scores? A: Fugu achieves high benchmark scores by intelligently orchestrating a diverse pool of specialized AI models. It dynamically selects the most appropriate agent for each subtask, allowing it to outperform monolithic models that try to handle everything themselves.
Q: Does Fugu Ultra beat Claude Fable 5? A: On LiveCodeBench, yes — Fugu Ultra (93.2) and even standard Fugu (92.9) beat Fable 5 (89.8). On SWE-Bench Pro, Fable 5 scores higher at 80.0 against Ultra's 73.7, but Fable 5 was pulled from public access on June 12, 2026 under U.S. export controls. All scores are vendor-reported.
Q: Can Fugu replace existing frontier models like GPT-5.5 or Fable 5? A: Fugu does not replace these models; rather, it coordinates them (or similar models in its agent pool). It acts as a conductor for an orchestra of AI agents, leveraging their individual strengths to achieve superior overall performance.
Q: How long does a Fugu Ultra request take? A: Complex tasks can take 15–25 minutes, because the orchestrator fans out to multiple experts, runs verification loops, and fuses results. For latency-sensitive work, use the standard tier.
Q: Why might my Fugu bill be higher than expected? A: Fugu Ultra bills orchestration tokens — the background tokens spent delegating sub-tasks, verifying code and routing between agents. They are charged at standard rates and can make the effective cost much higher than the visible output length suggests.
Q: Can I use Sakana Fugu with my existing OpenAI SDK code?
A: Yes. Fugu is served through an OpenAI-compatible API at https://api.sakana.ai/v1; change the base URL. It is also available via OpenRouter.
Q: Can I use Sakana Fugu in the EU or UK? A: No. Sakana does not provide Fugu services in EU or EEA member states while it works on GDPR compliance, and UK users have reported access issues. The OpenRouter path may work as an alternative in some cases.
Q: What are the main benefits of using a multi-agent orchestration system like Fugu? A: The main benefits include superior performance on complex, multi-step tasks, enhanced problem-solving capabilities, consistent persona stability in agentic applications, and the ability to achieve frontier-level AI without direct access to potentially restricted monolithic models.
Q: Where can I find more technical details about Sakana Fugu? A: Sakana AI has published a technical report on GitHub, and the approach builds on the ICLR 2026 papers TRINITY and Conductor.
Q: Is Sakana Fugu an open-source model? A: No, Fugu Ultra is not open source; the seed data lists it as proprietary. However, its architecture focuses on orchestrating other models, some of which may be open-source. The orchestration research is public.
Every claim here is traced to a primary source, dated, and listed under Sources. Research and drafting are AI-assisted; editing, verification and publication are human decisions, and a person is accountable for what appears on this page. How we work →
Discussion
0 comments