Fugu Ultra
Learned multi-agent orchestration system from Sakana AI that routes tasks across frontier models (Claude, Gemini, GPT) and can recursively call itself. Leads 6 of 11 benchmarks (vendor-reported) with a 1M-token context window, though performance depends on third-party API availability. Available via Sakana AI API or OpenRouter.
Specifications
| Lab | Sakana |
|---|---|
| Context window | 1M tokens |
| Input price | $5.00/1M |
| Output price | $30.00/1M |
| Release | Jun 2026 |
Strengths
- Top-tier multi-step reasoning and coding benchmarks
- 1-million-token context window for long documents
- Dynamic routing across multiple frontier models
- Built-in web search and configurable reasoning effort
Weaknesses
- Higher latency vs. base Fugu model
- Agent pool is fixed, not customizable
- Performance depends on third-party API uptime
- Orchestration tokens incur additional real costs
Best for
- Competitive coding and Kaggle challenges
- Complex multi-step AI research tasks
- Cybersecurity analysis and penetration testing
- Advanced agentic workflows with web search
In Depth: Fugu Ultra
Benchmark Performance
Fugu Ultra claims state-of-the-art on 6 of 11 evaluations, achieving 73.7 on SWE-Bench Pro and 93.2 on LiveCodeBench surpassing comparable frontier systems. However, these results are vendor-reported and independent verification has not yet been confirmed.
According to Sakana AI, Fugu Ultra tops Humanity's Last Exam at 50.0, GPQA-D at 95.5, and TerminalBench 2.1 at 82.1. It trails the base Fugu on SciCode and Long-Context Reasoning. These benchmarks are self-reported; users should treat them as preliminary until third-party validation emerges.
Pricing & Value
Input tokens cost $5 per million, output tokens $30 per million with higher rates for contexts above 272K tokens. This places Fugu Ultra at a premium over base Fugu but competitive given its multi-model orchestration.
At $30/M output tokens, Fugu Ultra's price-per-point on LiveCodeBench (93.2) is roughly $0.32 per point competitive with top-tier models like Claude Opus or GPT-5.5 on similar benchmarks. However, orchestration overhead can add hidden costs. Base Fugu offers lower latency and no orchestration surcharges.
Who Should Use This
Developers building agentic pipelines, AI researchers validating multi-step reasoning, and cybersecurity analysts needing deep code analysis. Not ideal for latency-sensitive or custom-agent deployments.
- Competitive coders: top-tier LiveCodeBench scores, but watch latency and orchestration costs
- AI researchers: leads Humanity's Last Exam, but vendor-reported benchmarks only
- Cybersecurity analysts: excels in terminal and code tasks, but agent pool is fixed
- Anti-persona: teams needing low-latency inference or custom model routing should consider base Fugu
Release & Version History
Launched June 22, 2026, Fugu Ultra builds on Sakana AI's multi-agent research published at ICLR 2026. It is a learned orchestration layer, not a traditional monolithic model.
Fugu Ultra follows the earlier base Fugu release, which offered lower latency and optional model opt-outs. Sakana AI released two ICLR 2026 papers and an arXiv technical report (2606.21228) detailing the system's recursive multi-agent architecture. Weights are not distributed; access is API-only via Sakana or OpenRouter. Regional restrictions may apply (unconfirmed EU/EEA limitation reported by third parties).