# Fugu Ultra

Sakana AI's quality-maximizing multi-agent system delivered as a single model API, coordinating a pool of expert agents within a context window of 1M tokens.

- **Model ID:** `sakana/fugu-ultra`
- **Type:** chat
- **Providers:** sakana
- **Context window:** 1,000,000
- **Maximum output tokens:** 1,000,000
- **Pricing:** $5/1M input tokens, $30/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/fugu-ultra

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'sakana/fugu-ultra',
  prompt: 'Why is the sky blue?'
})
```

## About

Fugu Ultra was released June 21, 2026 as Sakana AI's premium tier in the Fugu line. Instead of answering with a single network, Fugu Ultra runs a multi-agent system behind one model API. A lightweight coordinator assigns Thinker, Worker, and Verifier roles across a pool of frontier models, then routes between one and three agents depending on the problem. The approach builds on Sakana AI's published TRINITY and Conductor research.

Sakana AI tuned Fugu Ultra for maximum answer quality rather than low latency, drawing on a deeper, fixed agent pool and spending more compute per request. Early users applied Fugu Ultra to AI research, paper reproduction, cybersecurity analysis, and literature and patent investigations. In Sakana AI's published evaluations, Fugu Ultra scores 73.7 on SWE-Bench Pro, 95.5 on GPQA-Diamond, and 50.0 on Humanity's Last Exam.

Fugu Ultra supports a context window of 1M tokens with output up to 1M tokens, plus prompt caching for repeated context. Through AI Gateway, you call Fugu Ultra with a unified API key, automatic retries, and built-in observability, using the AI SDK, Chat Completions API, Responses API, Messages API, or other API formats.

## What to consider

Fugu Ultra optimizes for answer quality, not speed. Each request can fan out across multiple coordinated agents, so responses take longer than single-model calls. See live metrics on this page, and reserve Fugu Ultra for problems where quality justifies the wait.

Pricing is tiered by context length. Requests beyond 272K tokens bill at higher rates for input, output, and cached input. Track actual spend with AI Gateway's built-in observability, and trim prompts where you can.

The agent pool behind Fugu Ultra is fixed, and per-request routing details aren't exposed. Teams with strict model-provenance requirements should confirm the orchestrated approach fits their compliance needs before deploying. Fugu Ultra is available through Sakana AI.

## When to use

### Best For

- **Deep research workflows** including literature reviews, patent investigations, and paper reproduction that benefit from coordinated multi-agent analysis
- **Hard multi-step problems** in engineering and science where answer quality outweighs latency and cost per request
- **Cybersecurity analysis** requiring careful multi-step reasoning over code, configurations, and threat reports
- **Long-context investigations** leveraging the context window of 1M tokens for large codebases and document sets
- **Quality-critical agent steps** where you route the hardest parts of a pipeline to Fugu Ultra and compare listed rates ($5 input, $30 output per million tokens)

### Consider Alternatives When

- Latency matters more than peak answer quality. Single-model options such as GPT-5.4 or Gemini 3.5 Flash respond faster for interactive chat
- Your workload is high-volume classification, extraction, or summarization. Efficiency models such as DeepSeek V4 Flash or Kimi K2.6 cost less per request
- You must know exactly which model produced each answer. Fugu Ultra routes across a fixed agent pool without exposing per-request details
- Budget is the primary constraint. Premium per-token rates make Fugu Ultra expensive for routine generation at high volume

## Best for

- **Deep Research Workflows:** Literature reviews, patent investigations, and paper reproduction benefit from coordinated multi-agent analysis
- **Hard Multi-Step Problems:** Engineering and scientific questions where answer quality outweighs latency and cost per request
- **Cybersecurity Analysis:** Careful multi-step reasoning over code, configurations, and threat reports fits the verifier-backed workflow
- **Long-Context Investigations:** The context window of 1M tokens fits large codebases, document sets, and extended research threads
- **Quality-Critical Agent Steps:** Route your hardest pipeline steps to Fugu Ultra while cheaper models handle routine work

## Consider alternatives

- **Latency-Sensitive Applications:** Multi-agent coordination adds response time, so single-model options such as GPT-5.4 or Gemini 3.5 Flash fit interactive chat better
- **High-Volume Simple Tasks:** Classification, extraction, and summarization run cheaper on efficiency models such as DeepSeek V4 Flash or Kimi K2.6
- **Predictable Single-Model Behavior:** Teams that must know exactly which model produced each answer should pick a conventional model, since Fugu Ultra routes internally
- **Budget-Constrained Workloads:** Premium per-token rates make Fugu Ultra expensive for routine generation at high volume

## Frequently asked questions

### How is Fugu Ultra different from a conventional language model?

Fugu Ultra is a multi-agent system delivered as a single model API. A lightweight coordinator assigns Thinker, Worker, and Verifier roles across a pool of frontier models, routing between one and three agents per problem. You call one endpoint and receive one answer, but multiple coordinated models may work on the request.

### What is the context window for Fugu Ultra?

1M tokens, with output up to 1M tokens. That fits large codebases, long document sets, and extended research threads in a single request. Requests beyond 272K tokens bill at higher tiered rates.

### How much does Fugu Ultra cost through AI Gateway?

Current rates appear in the pricing panel on this page and update as providers adjust them. Pricing is tiered by context length, with higher rates for input, output, and cached input beyond 272K tokens.

### Does Fugu Ultra support Zero Data Retention?

Zero Data Retention is not currently available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

### How do I authenticate with Fugu Ultra through AI Gateway?

AI Gateway provides a unified API key. You don't need a separate Sakana AI account. Configure your API key, then use the model identifier to route requests through the AI SDK, Chat Completions API, Responses API, or Messages API. BYOK is also supported if you have a direct provider account.

### Why do responses from Fugu Ultra take longer than single-model calls?

Each request can involve multiple coordinated agents working in sequence, including verification passes. That extra orchestration improves answer quality on hard problems but adds latency. See live metrics on this page, and use faster single-model options for interactive workloads.

### Which underlying models does Fugu Ultra use to answer a request?

Sakana AI draws from a fixed pool of frontier models for Fugu Ultra and doesn't disclose per-request routing. If your compliance requirements demand knowing exactly which model handled each request, choose a conventional single model instead.

## Links

- [Model page](https://vercel.com/ai-gateway/models/fugu-ultra)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://console.sakana.ai/models)