Plain old default-HTML edition.
Bottom line: Sakana Fugu is not a normal single LLM. It is a learned orchestration layer presented as one model/API. You call fugu or fugu-ultra; internally it may route, delegate, verify, and synthesize work across a pool of frontier model agents.
Primary sources: Sakana Fugu product page, SakanaAI/fugu GitHub repo, Fugu technical report, TRINITY paper, and Conductor paper.
Sakana describes Fugu as a "multi-agent system as a model". The user sees one model interface, but behind that interface Fugu can select worker models, assign subtasks, coordinate communication, and synthesize a final answer. See the Fugu product page and GitHub README.
The public repo shows an OpenAI-compatible Sakana API provider at https://api.sakana.ai/v1, model slugs fugu and fugu-ultra, and a Codex launcher called codex-fugu. Sources: model config, provider config, command reference.
Fugu is the latency-aware version. It is designed for everyday interactive work and generally selects a worker model for the current input or interaction state. Source: Fugu technical report, section 3.1.
Fugu Ultra is the quality-first version. It builds deeper multi-agent workflows and is intended for difficult, multi-step tasks where quality matters more than latency. Source: product FAQ and technical report, section 3.2.
Fugu's job is to make collective model intelligence usable without making the user design an agent team. It can choose models, assign subtasks, verify outputs, and combine responses depending on the problem. Sakana positions this as a way to reduce single-vendor dependency and combine complementary model strengths. Source: product page.
Reported use cases include coding, code review, agentic software engineering, scientific reasoning, autonomous ML experimentation, paper reproduction, literature and patent analysis, cybersecurity analysis, classical Japanese letter reading-order recovery, and CAD generation. Source: technical report, section 4.
The lower-latency Fugu variant builds on the TRINITY idea. A compact coordinator reads the task/context, uses hidden-state representations, and a lightweight head scores/selects a worker model. In the original TRINITY paper, the coordinator also assigns roles such as Thinker, Worker, and Verifier. The production Fugu variant simplifies the latency-sensitive path by selecting a worker without role assignment. Sources: TRINITY paper, technical report, section 3.1.
Training for Fugu is described in two stages: supervised fine-tuning on measured single-step worker performance, followed by evolutionary optimization on end-to-end multi-turn tasks using sep-CMA-ES. Source: technical report, sections 3.1.2-3.1.3.
Fugu Ultra builds on the Conductor framework. The Conductor is an LLM trained to output a natural-language workflow: subtasks, assigned worker agents, and access lists controlling which previous outputs each worker can see. This allows chains, trees, debate, independent attempts, specialist review, and final aggregation. Sources: Conductor paper, technical report, section 3.2.
The Conductor is trained with reinforcement learning/GRPO. Its reward checks whether the workflow format is parseable and whether executing the workflow produces a correct final answer. Source: Conductor paper, section 3.
Fugu Ultra's production design must preserve agent identity across tool/function calls. The report describes workflow state tracking, intra-workflow agent isolation to avoid "orchestration collapse," and persistent shared memory across workflows so agents do not repeatedly rediscover the same context. Source: technical report, section 3.2.2.
Sakana reports that Fugu and Fugu Ultra outperform or match public frontier baselines across many coding, reasoning, scientific, and agentic benchmarks. Caveat: these are Sakana/vendor-reported results, and several baseline scores are provider-reported. Source: product page quantitative table, technical report, table 1.
| Benchmark | Fugu | Fugu Ultra | Notable comparison |
|---|---|---|---|
| SWE Bench Pro | 59.0 | 73.7 | Opus 4.8: 69.2; GPT 5.5: 58.6 |
| TerminalBench 2.1 | 80.2 | 82.1 | GPT 5.5: 78.2 |
| LiveCodeBench | 92.9 | 93.2 | Gemini 3.1 Pro: 88.5 |
| LiveCodeBench Pro | 87.8 | 90.8 | GPT 5.5: 88.4 |
| Humanity's Last Exam | 47.2 | 50.0 | Opus 4.8: 49.8 |
| CharXiv Reasoning | 85.1 | 86.6 | Opus 4.8: 84.2 |
| GPQA-Diamond | 95.5 | 95.5 | Gemini 3.1 Pro: 94.3 |
| SciCode | 60.1 | 58.7 | Gemini 3.1 Pro: 58.9 |
| Long Context Reasoning | 74.7 | 73.3 | GPT 5.5: 74.3 |
| MRCRv2 | 86.6 | 93.6 | GPT 5.5: 94.8 |
Not every metric is a win. For example, GPT 5.5 is reported higher than Fugu Ultra on MRCRv2. The stronger claim is not universal dominance, but that learned orchestration gives large gains on many hard multi-step and agentic tasks.
The product page lists subscriptions at Standard $20/month, Pro $100/month, and Max $200/month. It also describes pay-as-you-go pricing for enterprise/heavy workloads. Source: pricing section.
For Fugu, when one agent is active, Sakana says you pay the standard rate for that underlying model. When multiple agents are active, it says fees are not stacked; instead, you pay a single rate based on the top-tier model involved. Source: pricing FAQ.
For Fugu Ultra model fugu-ultra-20260615, the page lists per-1M-token rates: $5 input, $30 output, $0.50 cached input; above 272K context, $10 input, $45 output, $1 cached input. Source: pricing section.
Availability caveat: Sakana says Fugu is not yet available in the EU/EEA while it works toward GDPR and EU-specific compliance. Source: product page notice and FAQ.
Fugu is best understood as a productionized learned model-router/conductor, not a newly trained monolithic foundation model. Its bet is that frontier capability can be improved by behavior-level composition: route, coordinate, verify, and synthesize across models rather than training one larger model or merging weights.
For simple calls, a single cheap/fast model may still be the right tool. For hard tasks involving debugging, long-horizon agentic work, research, or mixed-domain reasoning, Fugu Ultra's orchestration is the interesting part.