Sakana Fugu Ultra is a multi-agent orchestration system from Tokyo-based Sakana AI that behaves like a single AI model from the outside while coordinating a pool of frontier LLMs on the inside. On LiveCodeBench v6, Fugu Ultra scores 93.2—beating every publicly accessible standalone model including Claude Fable 5 (89.8)—and it costs $5 per million input tokens and $30 per million output tokens on pay-as-you-go pricing. The core idea: instead of scaling one massive model, train a small orchestrator model to decompose tasks, delegate to specialized expert models, verify their work, and fuse the results into one answer.
Last verified: 2026-08-01 · Multi-agent orchestration, not a single monolith · Two tiers: Fugu (fast) and Fugu Ultra (deep) · EU/EEA restricted · Vendor-reported benchmarks—test on your workloads
How Does Sakana Fugu Work?
Sakana Fugu is itself a small language model trained specifically to orchestrate other models. When you send a request to its single OpenAI-compatible API endpoint, the Fugu model decides whether to answer directly or assemble a team of expert models from its internal pool. It handles model selection, task delegation, verification, and synthesis internally—so from your code's perspective, it looks like calling one model. The architecture is built on two ICLR 2026 papers from Sakana AI: TRINITY (an evolved LLM coordinator) and the Conductor (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. Source: sakana.ai/fugu-release
What Is the Difference Between Fugu and Fugu Ultra?
Fugu is the lower-latency default optimized for everyday coding, code review, chat, and interactive work. Fugu Ultra coordinates a deeper pool of expert agents and is tuned for maximum answer quality on hard, multi-step problems like AI research, cybersecurity analysis, and multi-step engineering tasks. Both are accessible through the same API. The key difference for developers: Fugu lets you opt out of specific models from the console's settings menu for data privacy or compliance, while Fugu Ultra's pool is fixed because it relies on the full agent set to deliver its benchmark performance. Source: sakana.ai/fugu
| Feature | Fugu | Fugu Ultra |
|---|---|---|
| Optimized 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 top model used) | Fixed $5/$30 per 1M tokens |
| Best for | Coding, chat, code review | Research, security, long-horizon tasks |
How Fast Is Sakana Fugu Ultra?
Fugu Ultra is not fast. A complex task that fans out to multiple expert models and runs verification loops can take 15–25 minutes of orchestrated thinking. The tradeoff is intentional: Fugu Ultra prioritizes answer quality over latency for hard, multi-step problems where a single model call often produces 90% correct output with 10% broken. For everyday coding or chat, the standard Fugu tier is the better choice—it balances quality with low latency. The practical approach: run Fugu Ultra in the background for one-off, high-stakes tasks while continuing other work on a faster model. Source: sakana.ai/fugu-beta
How Much Does Sakana Fugu Cost?
Sakana Fugu offers both subscription plans and pay-as-you-go pricing. 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 costs $5 per million input tokens and $30 per million output tokens ($0.50 for cached input), with rates doubling to $10/$45 per million tokens when 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. Source: console.sakana.ai/pricing
| Pricing tier | Input (per 1M) | Output (per 1M) | Cached input | Context > 272K |
|---|---|---|---|---|
| Fugu Ultra standard | $5.00 | $30.00 | $0.50 | $10/$45/$1.00 |
| Fugu Ultra > 272K ctx | $10.00 | $45.00 | $1.00 | — |
| Fugu Cyber | $6.00 | $36.00 | $0.60 | $12/$54/$1.20 |
Important caveat: Fugu Ultra bills orchestration tokens. The background tokens consumed when Fugu delegates sub-tasks, verifies code, and routes between agents are counted toward your request at standard rates—they are not absorbed by the provider. A task with heavy fan-out and verification can cost far more than its visible output length suggests. Source: console.sakana.ai/pricing
How does Fugu Ultra's cost compare to other frontier models?
| 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 |
At the headline rate, Fugu Ultra matches GPT-5.5 and costs half of what the restricted Claude Fable 5 charges. But the effective rate depends on how many orchestration tokens a given task burns. Source: sakanafugu.novcog.us.com/pricing
What Are the Sakana Fugu Benchmark Scores?
Sakana Fugu Ultra benchmarks are vendor-reported (Sakana AI) and have not yet been independently reproduced by third-party labs as of August 2026. Here is the full scoreboard from Sakana's published data:
| 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 |
| 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 |
The pattern: Fugu Ultra wins on decomposable, multi-step, agentic tasks (LiveCodeBench, SWE-Bench Pro, GPQA-Diamond) where combined model strengths matter. It 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 due to U.S. export controls. Source: sakanafugu.novcog.us.com/benchmarks, DataCamp
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, or the physics works but part of the application never renders. We have all been there: a long build comes back 90% right and 10% broken.
Multi-agent orchestration attacks this differently. Instead of one model doing everything, an orchestrator model decomposes the task, routes each sub-task to a specialized expert, and fuses the results. If one expert makes a mistake, another covers for it. The approach mirrors how human teams work—a senior architect breaks down a project, assigns pieces to specialists, reviews the work, and synthesizes the final product.
The key insight from Sakana's research is that the orchestrator itself is a trained model, not a prompt-engineering script. It learned through reinforcement learning when to delegate versus solve directly, how agents should communicate, and how to combine outputs. This is why Fugu is not just a model router like OpenRouter (which forwards requests to a model of your choice)—it is a trained coordinator that adapts its strategy per task. Source: sakana.ai/fugu-release
For builders working with AI agent operating systems, Fugu's architecture offers a compelling alternative to manually wiring multi-model pipelines. Instead of building custom orchestration with tools like Macaron-V1's mixture-of-LoRA approach, Fugu packages the orchestration as the model itself. Teams focused on serving AI agents at scale will recognize the pattern: offloading coordination to a trained model simplifies the deployment surface.
How Do You Access Sakana Fugu?
There are two primary access paths:
Direct from Sakana AI — Visit console.sakana.ai to create an API key. The API is OpenAI-compatible at
https://api.sakana.ai/v1, so you can use existing OpenAI SDK code with a base URL change. Available in most regions globally.Through OpenRouter — If direct access is unavailable in your region, Fugu Ultra is also available via OpenRouter at the same $5/$30 per 1M tokens rate. This is the recommended path for users where direct access is blocked.
You can also run an alternative multi-model fusion approach using OpenRouter's Fusion API, which fans out one prompt to multiple models and uses a judge model to synthesize the answer—similar in concept but without Fugu's trained orchestrator.
Is Sakana Fugu Available in the EU?
No. Sakana explicitly does not provide Fugu services to users in EU (European Union) or EEA (European Economic Area) member states while it works to align its black-box routing architecture with GDPR compliance requirements. Users in the UK have also reported access issues. In other regions, access may be limited by network conditions or local regulations. If you are in a blocked region, the OpenRouter path may work as an alternative. Source: sakana.ai/fugu
When Should You Use Fugu Ultra vs a Single Model?
Fugu Ultra makes the most sense when orchestration can beat one cheaper model. It is ideal for:
- Long-horizon research — paper reproduction, patent searches, multi-source analysis where sustained progress across many steps matters
- Difficult code review — complex refactoring, security analysis, and architecture audit where multiple expert perspectives catch more issues
- Multi-step engineering — when a single model call produces 90% correct output with 10% broken across different dimensions (logic, UI, physics, controls)
- High-stakes one-shot tasks — where getting it right the first time is worth 25 minutes of orchestrated thinking
Fugu Ultra is not the right choice for:
- Simple coding or chat (use standard Fugu instead)
- Budget-sensitive workloads (DeepSeek-V4-Flash at $0.42/M combined will serve most routine tasks)
- Tasks requiring tight latency (the multi-agent fan-out adds minutes, not seconds)
- EU/EEA-based teams (access is currently blocked)
For teams experienced with automated AI research workflows or building revenue loops with AI agents, Fugu Ultra fits naturally as the high-quality layer for complex reasoning tasks.
What Are the Alternatives to Sakana Fugu?
If Fugu is not available, too expensive, or too slow for your use case, several alternatives achieve similar multi-model orchestration:
| Alternative | How it works | Cost | Best for |
|---|---|---|---|
| OpenRouter Fusion | Fans out one prompt to N models, uses a judge model to synthesize | Pay per model + judge | Budget-conscious teams wanting multi-model fusion |
| Fusion Engine (open source) | Self-hosted OpenRouter Fusion alternative with custom judge prompts | Per-model API costs only | Teams wanting full control over panel and judge |
| DIY orchestration (LangGraph/AutoGen) | Build your own routing, delegation, and synthesis | Engineering time + API costs | Teams with specific workflow needs |
| Standard Fugu | Same orchestrator, lighter agent pool | Dynamic (top model rate) | Everyday coding and chat |
The DIY approach is increasingly viable now that Sakana has published its orchestration research (TRINITY, Conductor) as ICLR 2026 papers. For teams already managing AI token cost optimization, building your own orchestration layer with open tools may be more cost-effective at scale.
What This Means for You
The people winning with AI in 2026 are not picking one model and sticking with it. They are running a stack and routing each job to whichever brain—or team of brains—fits it. Sakana Fugu packages that routing intelligence into a single API call. For one-off, high-stakes tasks where quality matters more than speed or cost, Fugu Ultra is the strongest publicly accessible option. For everyday work, standard Fugu or a cheaper single model will serve you fine. The mindset shift: own the whole aquarium, not just your favorite fish.
FAQ
Q: Is Sakana Fugu a normal AI model? A: No. Fugu is a multi-agent orchestration system that behaves like a single model. It is itself a small language model trained to coordinate a pool of other frontier LLMs—handling model selection, delegation, verification, and synthesis behind one API endpoint.
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 versus Fugu Ultra's 73.7, but Fable 5 was pulled from public access on June 12, 2026 due to U.S. export controls. All scores are vendor-reported and not yet independently verified.
Q: How long does a Fugu Ultra request take? A: Complex tasks can take 15–25 minutes because the orchestrator fans out to multiple expert models, runs verification loops, and fuses results. For latency-sensitive work, use the standard Fugu tier instead.
Q: Can I use Sakana Fugu with my existing OpenAI SDK code?
A: Yes. Fugu is available through an OpenAI-compatible API at https://api.sakana.ai/v1. You can use existing OpenAI SDK code with a base URL change. It is also available via OpenRouter.
Q: Why might my Fugu bill be higher than expected? A: Fugu Ultra bills orchestration tokens—the background tokens consumed when delegating sub-tasks, verifying code, and routing between agents. These are counted at standard rates and can make effective costs higher than the visible output length suggests.
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. UK users have also reported access issues. The OpenRouter path may work as an alternative in some cases.
Discussion
0 comments