Fugu Ultra for Law Firms

Fugu Ultra is a multi-agent orchestration system released by Sakana AI, a Tokyo-based AI company, on June 22, 2026. Rather than running its own model weights, it routes each part of a task to different frontier models - such as Claude, GPT, and Gemini - through a single OpenAI-compatible API, then returns one combined answer. Its learned coordination approach draws on two ICLR 2026 papers, TRINITY and Conductor.

Fugu Ultra is the higher-performance member of the Fugu family. The standard Fugu tier prioritizes lower latency and cost, while Ultra aims for stronger results on harder tasks. Fugu Ultra is available only as a hosted API; its weights are not open-source. Sakana lists usage pricing of $5 per million input tokens and $30 per million output tokens, alongside subscription tiers.

This page describes where Fugu Ultra can assist a law firm - drafting, summarizing, and comparison work across common workflows - and the practical limits that come with it. The headline risk for legal users is that your prompts leave Sakana and reach third-party model vendors, so client confidentiality depends on more than one company's terms.

What Fugu Ultra is, in plain terms

Fugu Ultra is not a single language model. It is an orchestrator that receives your request through one API and decides how to split the work across several frontier models operated by other companies. Each model handles the part it is suited to, and Sakana's system assembles the pieces into one response.

For a firm, the practical effect is that you interact with one endpoint but your text may be processed by Anthropic, OpenAI, or Google infrastructure depending on how Fugu Ultra routes the task. That single-API convenience is the appeal; the multi-vendor routing is also the source of the confidentiality question covered below.

• Single OpenAI-compatible API, multiple underlying frontier models.
• Hosted service only - no self-hosting, no on-premise weights.
• Ultra targets harder tasks; standard Fugu targets lower latency and cost.
• Released June 22, 2026 by Sakana AI (Tokyo).

Where it can help across firm workflows

As a general assistant, Fugu Ultra can support routine knowledge work that a firm already does by hand. Treat every output as a first draft that a lawyer reviews, not a finished work product.

The tasks below are ones where a capable general model can save time, provided a human checks the result against source documents and applicable law.

• Drafting first-pass internal memos, client update emails, and summaries.
• Summarizing long documents into short briefings for review.
• Comparing two documents or versions and flagging differences to check.
• Reformatting or reorganizing text you already have into a cleaner structure.
• Generating question lists or outlines to guide human research.

Confidentiality and privilege: the routing question

This is the central legal issue with Fugu Ultra. Because it routes your prompts to third-party frontier vendors - Anthropic, OpenAI, or Google - client confidentiality and privilege depend on both Sakana's data-handling terms and the terms of every downstream vendor in the chain. Your text does not stay with one provider.

Sakana has not stated that it offers a Business Associate Agreement (BAA) or any specific certification for Fugu. Do not assume one exists. Before entering any client-identifying information, a firm should confirm in writing how prompts are stored, logged, and routed by every model that may handle the request, including whether any provider retains data or uses it for training.

Under ABA Model Rule 1.6, lawyers must make reasonable efforts to prevent unauthorized disclosure of client information. Sending privileged material through a multi-vendor pipeline you have not vetted can put that duty at risk. Many firms address this by using only de-identified inputs, or by not using the tool for privileged matters until the data terms are confirmed.

Accuracy limits and required human review

Like all current AI assistants, Fugu Ultra can produce output that sounds confident but is wrong, including fabricated citations, invented quotations, and misstated holdings. This is not a hypothetical concern in law: in Mata v. Avianca, a federal court sanctioned attorneys who filed a brief containing AI-generated case citations that did not exist.

Every factual and legal claim in an AI output must be independently verified against primary sources before it is relied on or filed. Orchestration across multiple models does not remove this risk; a wrong answer from any model in the chain can still surface in the final response.

Firms should treat Fugu Ultra output as a starting point that a responsible attorney reviews, corrects, and takes ownership of under ABA Model Rules 5.1 and 5.3, which cover supervision of lawyers and non-lawyer assistance, including technology.

• Verify every citation, quotation, and holding against the primary source.
• Never file AI-generated text without attorney review (see Mata v. Avianca).
• Assign a responsible attorney to own the reviewed work product.

Honest fit: when to use it and when not to

Fugu Ultra is a reasonable fit for low-stakes, review-heavy drafting and summarizing where inputs can be de-identified and a lawyer checks the output. It is a poor fit for any workflow that requires entering privileged or client-identifying data before the routing and retention terms are confirmed in writing.

It is also a poor fit for firms that need a signed BAA, a specific compliance certification, or single-vendor data residency today, because Sakana has not stated those for Fugu. Cost-sensitive, high-volume, low-complexity tasks may be better served by the standard lower-cost Fugu tier or a single model, since Ultra carries higher output-token pricing at $30 per million output tokens.

• Good fit: de-identified drafting, summarizing, and comparison with human review.
• Poor fit: privileged data before written routing/retention confirmation.
• Poor fit: requirements for a BAA or specific certification Sakana has not stated.
• Consider standard Fugu or a single model for high-volume, low-complexity work.