Sakana Fugu Review — Multi-Agent Orchestration Packaged as a Single Model
Sakana Fugu Review — Multi-Agent Orchestration Packaged as a Single Model
📖 What Is Sakana Fugu?
Sakana Fugu is a multi-agent orchestration system that behaves like a single model. You call one OpenAI-compatible API, and behind the scenes, a learned orchestrator routes your query across a pool of frontier LLMs — Claude, GPT, Gemini, and others — to produce the best answer. It's not a framework you configure. It's a model that learned how to run a multi-agent system.
Launched June 22, 2026 by Tokyo-based Sakana AI, Fugu comes in two variants: Fugu (the latency-balanced version for everyday use) and Fugu Ultra (the quality-first version for complex, multi-step tasks). The orchestrator itself is a small (~7B parameter) language model trained using the TRINITY and Conductor architectures (both ICLR 2026 papers) to adaptively build agentic scaffolds per query [1].
The core idea is subtle but powerful: instead of you writing agent pipelines, routing logic, and model fallbacks, a learned controller does it dynamically. The orchestrator decides when to decompose a query into sub-tasks, which model to route each sub-task to, whether to verify results, and when to re-query with refined prompts. The user just sends a prompt and gets a response — the multi-agent complexity is hidden.
Sakana AI describes Fugu as a "collective intelligence" system: no single frontier model is best at everything, so a system that can pick the right expert per sub-task outperforms any individual model on aggregate. The benchmark results support this — Fugu Ultra matches or beats Claude Fable 5, Opus 4.8, and GPT-5.5 across coding, reasoning, and scientific benchmarks [1][2].
📊 At a Glance & ✅ Pros & Cons
| Feature | Sakana Fugu (Base) | Fugu Ultra | Claude Fable 5 | GPT-5.5 |
|---|---|---|---|---|
| Category | Multi-Agent Orchestration | Multi-Agent Orchestration | LLM | LLM |
| Architecture | ~7B controller + frontier pool | ~7B controller + frontier pool | Monolithic frontier LLM | Monolithic frontier LLM |
| LiveCodeBench | 89.4% | 93.2% [1] | ~92% | ~88% |
| SWE-Bench Pro | 65.1% | 73.7% [2] | — | 58.6% |
| GPQA Diamond | 91.2% | 95.5% [1] | ~94% | ~90% |
| API Format | OpenAI-compatible | OpenAI-compatible | Anthropic API | OpenAI API |
| Streaming | ✅ Yes | ❌ No | ✅ Yes | ✅ Yes |
| Pricing | $5/1M tokens | $30/1M tokens | $15/1M tokens | $10/1M tokens |
| Vendor Lock-in | Multi-model (none) | Multi-model (none) | Anthropic only | OpenAI only |
✅ What It Does Best
- Frontier performance without vendor lock-in — Fugu Ultra matches or beats Claude Fable 5 and GPT-5.5 on coding and reasoning benchmarks (93.2% LiveCodeBench, 73.7% SWE-Bench Pro) by dynamically routing to the best model per subtask [1].
- Single API, zero orchestration code — drop-in replacement for any OpenAI-compatible client. No agent framework setup, no manual routing logic, no prompt engineering per model.
- Learned orchestration beats hardcoded pipelines — the ~7B controller adapts its scaffolding per query using TRINITY/Conductor architecture, outperforming CrewAI and LangGraph on multi-step tasks [2].
- Two tiers for cost flexibility — base Fugu for latency-sensitive work, Fugu Ultra for max quality. You pick the tier, not the model.
- Transparent usage tracking — API returns per-request orchestration cost breakdown so you know exactly which models were used and at what cost.
❌ Where It Falls Short
- High latency on Ultra tier — Fugu Ultra can take 30+ minutes for complex multi-step queries as the orchestrator iterates and verifies. Not suitable for interactive use [4].
- Cost is unpredictable — orchestration tokens and multi-model routing mean a single query can cost $5-10 on Ultra. No hard budget caps at launch.
- Black-box routing — you don't control which models the orchestrator picks. No override mechanism or model whitelist for compliance-sensitive workloads.
- Early-stage support — Sakana is a small startup (Tokyo-based). Community is nascent, docs are thin beyond the technical report, and enterprise SLAs aren't published.
See all multi-agent orchestration tools compared side-by-side in the ToolBrain database.
Popular open-source multi-agent framework with manual workflow configuration. More control, more work. 7.8/10 on ToolBrain.
LangChain's graph-based agent orchestration. Powerful for complex pipelines, steep learning curve. 7.8/10 on ToolBrain.
Anthropic's top-tier frontier model. Single-model simplicity, but vendor-locked to Anthropic's ecosystem. 8.0/10 on ToolBrain.
🎯 Score Breakdown
🦾 Ease of Use
Fugu's big selling point: you don't set up multi-agent infrastructure. Point your OpenAI-compatible client at api.sakana.ai and you're done. The API surface is identical to what you already use — same chat completion format, same tool calling, same streaming on base Fugu. There's nothing to learn beyond picking the right model ID. The complexity lives entirely on Sakana's side. That said, understanding when to use Fugu vs a direct model API takes experience. If you're paying $30/M tokens on Ultra for simple summarization, you're wasting money. The platform doesn't guide you on cost-optimal usage patterns yet — you learn by monitoring the cost breakdown on each response. For teams new to multi-agent systems, Fugu's simplicity is a genuine win. For cost-conscious teams, the lack of guardrails is friction.
⚙️ Features
Fugu packs an extraordinary amount of capability into a simple API. Dynamic multi-model routing, automatic task decomposition, iterative verification, self-correction loops, tool/function calling support, streaming on base tier, structured output (JSON mode), usage cost breakdown per request, and a growing pool of backend models that updates as new frontier models release. The technical report describes the orchestrator's ability to craft agentic scaffolds on the fly — spinning up specialized sub-agents for research, coding, verification, and synthesis without any user configuration. That's a feature category that existing frameworks require hours to set up. The main feature gap is the lack of routing overrides: you can't specify "only use model X for this query" or block certain providers. For enterprise compliance teams, that's a blocker. But as a pure capability play, Fugu's feature set is remarkable for a v1 product.
🚀 Performance
The benchmarks are the headline. Fugu Ultra scores 93.2% on LiveCodeBench (ahead of Claude Fable 5), 95.5% on GPQA Diamond, and 73.7% on SWE-Bench Pro — significantly ahead of Opus 4.8 (69.2%) and GPT-5.5 (58.6%) [1][2]. The multi-model approach genuinely outperforms individual frontier models on aggregate because the orchestrator picks the best model for each sub-task. Base Fugu trades some of that ceiling for latency, hitting 89.4% on LiveCodeBench and 65.1% on SWE-Bench Pro — still competitive with frontier models. Fugu also scored 82.1% on TerminalBench 2.1 (agentic task execution) and strong results on Humanity's Last Exam and CharXiv Reasoning [2]. The performance caveat: benchmarks are Sakana's own reporting, and independent verification is still emerging. Latency is the real tradeoff — base Fugu is comparable to GPT-5.5 on response time, but Ultra can take 30+ minutes on complex queries [4]. For interactive use, stick with base. For deep research where quality matters more than speed, Ultra delivers.
📚 Documentation
Sakana released an excellent technical report (26 pages, arXiv: 2606.21228) that explains the architecture, training methodology, and benchmark methodology thoroughly. The TRINITY and Conductor papers (both ICLR 2026) provide deep academic grounding. But practical API documentation is thin. The console.sakana.ai site has basic endpoint docs, model IDs, and pricing — it covers the minimum to integrate. What's missing: example implementations beyond basic chat completion, error handling guides, rate limit documentation, timeout guidance for Ultra queries, cost optimization patterns, and integration examples with popular frameworks. The gap between the (excellent) research paper and the (sparse) API docs is roughly where most developers will live. For a v1 launch, this is forgivable. For production adoption, it needs significant investment.
🎯 Support
Sakana AI is a research lab first, a product company second. The GitHub repository has basic issues and discussions, but the community is days old (the product launched June 22). There's no Discord, no forum, no Stack Overflow presence. Support is email-only with no published SLAs. Enterprise support, dedicated account management, and uptime guarantees are not yet available. For a tool that could route your production workloads, this is the weakest area. The transparency of the published research partially compensates — you can understand exactly how the system works from the papers — but if something breaks, you're reliant on email to a small team in Tokyo. Early adopters should plan for self-service troubleshooting and have fallback model APIs ready.
🔬 How It Works: Orchestration as a Learned Skill
Fugu's architecture is the most interesting part. Most multi-agent frameworks (CrewAI, LangGraph, OpenAI Agents SDK) are configured — you write code that defines agents, tools, workflows, and handoff logic. Fugu is learned — a ~7B language model was trained to understand queries and dynamically devise agentic scaffolds to solve them.
The orchestrator model does three things per query [2]:
- Task decomposition — analyzes the query and decides if it should be handled directly or broken into sub-tasks. If decomposition is needed, it designs the sub-task structure.
- Model routing — for each sub-task, picks the best frontier model from the pool. The pool includes Claude (Sonnet, Opus, Fable 5), GPT (4o, 5.5), Gemini (Pro, Ultra), and others. The routing decision is based on the model's demonstrated strengths for the task type.
- Verification & iteration — optionally runs verification loops, checking outputs for correctness and re-querying if results fall below a confidence threshold. Fugu Ultra does this aggressively (hence the 30+ minute latency).
This learned approach means Fugu improves over time as new models enter the pool and the orchestrator is retrained. Sakana has committed to continuously updating the model pool and retraining coordinators to maintain Fugu's performance edge [1].
For developers, the practical upshot is: a single API call can trigger what would be a multi-agent CrewAI pipeline with 50+ lines of orchestration code. You lose visibility and control, but you gain the orchestrator's learned optimization — which, based on the benchmarks, is already outperforming most manual configurations.
📊 Benchmark Deep Dive: Fugu Ultra vs The Field
| Benchmark | Fugu Ultra | Fugu (Base) | Claude Fable 5 | Claude Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|---|---|
| LiveCodeBench | 93.2% | 89.4% | ~92% | ~85% | ~88% | ~82% |
| SWE-Bench Pro | 73.7% | 65.1% | — | 69.2% | 58.6% | 54.2% |
| GPQA Diamond | 95.5% | 91.2% | ~94% | ~89% | ~90% | ~87% |
| TerminalBench 2.1 | 82.1% | 74.6% | — | — | — | — |
| Humanity's Last Exam | 18.5% | 14.2% | ~17% | ~12% | ~15% | ~11% |
Sources: Sakana Fugu Technical Report [1], VentureBeat [2], and Sakana's benchmark dashboard [3]. Claude Fable 5 benchmarks are approximate as it was temporarily suspended during testing.
🎯 Ideal Use Cases
✅ Best For
|
❌ Not Ideal For
|
Free tier: limited evaluation requests with basic rate limits. Pro ($20/mo): base Fugu access, 500K tokens/day, standard queue. Max ($200/mo): both tiers, 5M tokens/day, priority queue. Pay-as-you-go: $5/1M tokens (base Fugu), $30/1M tokens (Fugu Ultra), no subscription required [3].
Quick start: Sign up at console.sakana.ai → get an API key → point your OpenAI client at https://api.sakana.ai/v1 with model ID fugu-20260615 or fugu-ultra-20260615 → send your first request.
| ❓ FAQ | |
|---|---|
| What is Sakana Fugu? | Sakana Fugu is a multi-agent orchestration system that routes queries across frontier LLMs (Claude, GPT, Gemini, etc.) through a single OpenAI-compatible API. A learned ~7B orchestrator model handles task decomposition, model routing, and verification automatically. |
| How is Fugu different from CrewAI or LangGraph? | CrewAI and LangGraph are frameworks where you write orchestration code. Fugu compresses all of that into a learned controller model. You don't design the pipeline; the orchestrator learns to build one dynamically per query. Less control, but zero configuration. |
| Does Fugu support streaming? | Base Fugu supports streaming. Fugu Ultra does not — it runs iterative verification loops that require full context before returning results. |
| Is Sakana Fugu open source? | The orchestrator model weights are on GitHub under a research license, but the multi-model routing backend requires the hosted API. You can run the controller locally but not the full orchestration. |
| What benchmarks does Fugu lead? | Fugu Ultra: 93.2% LiveCodeBench, 95.5% GPQA Diamond, 73.7% SWE-Bench Pro, 82.1% TerminalBench 2.1 [1][2]. Full data in the technical report (arXiv: 2606.21228). |
| Can I control which models Fugu uses? | Not yet. Routing is fully automated with no override mechanism. Sakana has indicated model whitelisting is on the roadmap but not available at launch. |
| 📖 Related Reads | |
|---|---|
| CrewAI Review 2026 | Manual multi-agent orchestration framework. More control, more code. 7.8/10. |
| LangGraph Review 2026 | Graph-based agent orchestration from LangChain. Powerful for complex pipelines. 7.8/10. |
| OpenAI Agents SDK Review 2026 | OpenAI's agent framework with handoffs and guardrails. 7.8/10. |
| OpenAI Symphony Review 2026 | Multi-agent orchestration at scale. 7.5/10. |
