Sakana: Fugu Ultra v2

Fugu Ultra v2 is Sakana AI's peak-performance offering in the Fugu family: a learned multi-agent orchestration system exposed as a single OpenAI-compatible chat model (API ID fugu-ultra-v2.0). Instead of routing every query through one monolithic foundation model, a dedicated orchestrator language model plans dynamic agentic scaffolds, selects workers from a fixed pool of open and specialized models, and can recursively invoke itself to decompose hard problems. The v2 release emphasizes answer quality on complex multi-step reasoning, autonomous research, and full-stack software development while deliberately excluding certain proprietary frontier models (e.g., Fable 5/5.1 and GPT-6-Astra) from its pool.

On Sakana's September 2026 evaluation suite, Fugu Ultra v2 reports leading or tied-first results on five of eight agentic benchmarks—including DeepSWE (74.3) and Toolathon (80.6)—and ranks in the top two on seven of eight. It is strong on graduate-level science (GPQA-Diamond 95.5) and Humanity's Last Exam text (56.3), and supports configurable reasoning effort, function calling, structured JSON outputs, built-in web search, and multimodal inputs (text, images, and PDFs) over a one-million-token context window.

Fugu Ultra v2 is API-only: there are no public weights to self-host. Pricing is positioned as a premium tier relative to Fugu Max ($5/M input and $30/M output tokens at standard context, with higher rates above 272K tokens). Knowledge cutoff is August 28, 2026.