Sakana Fugu Ultra

Sakana AI's learned multi-agent orchestration system (GA June 22, 2026), delivered as a single OpenAI-compatible model. Fugu dynamically routes and delegates across a swappable pool of frontier LLMs (GPT/Claude/Gemini) plus recursive self-calls using Thinker/Worker/Verifier roles; the Ultra tier targets hard multi-step reasoning, coding and science. NOTE: it is an orchestration layer over third-party frontier models, not a single trained model. Cached input ~$0.50/1M; prices rise above 272K context (~$10 in / $45 out). Available via the Sakana API, console.sakana.ai and OpenRouter; not available in the EU/EEA. Benchmarks vendor-reported — verify.

Sakana Fugu Ultra strengths

• Learned multi-agent orchestration via one OpenAI-compatible API
• Routes across a swappable pool of frontier models + recursive self-calls
• Thinker/Worker/Verifier roles for hard multi-step tasks
• Vendor-agnostic resilience (reduces model lock-in)
• Per-request token/cost reporting

Pricing & context

Cost guide: a typical call of ~10K input + 2K output tokens runs roughly $5.00 × 0.01 + $30.00 × 0.002 — worth modelling against cheaper tiers before committing high-volume traffic.

When to choose Sakana Fugu Ultra

Sakana Fugu Ultra is best for Hard multi-step reasoning, coding and code review, scientific/research tasks, and quality-critical agentic workflows. If your workload is more cost-sensitive, weigh it against Llama 4 Scout (~$0.08 (varies by host) input /1M) first.

Sakana Fugu Ultra FAQ

How much does Sakana Fugu Ultra cost?

Sakana Fugu Ultra is priced at $5.00 per 1M input tokens and $30.00 per 1M output tokens (public API list price), with a 1M tokens (128K max output; tier changes above 272K) context window.

What is Sakana Fugu Ultra best for?

Sakana Fugu Ultra by Sakana AI is best for Hard multi-step reasoning, coding and code review, scientific/research tasks, and quality-critical agentic workflows.

Is Sakana Fugu Ultra multimodal?

Sakana Fugu Ultra supports text, image.