Fugu Ultra for Legal Research

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

Ultra is the higher-performance tier of the Fugu family, aimed at harder tasks, while standard Fugu prioritizes lower latency and cost. Fugu Ultra is a hosted API only; its weights are not open-source. Sakana lists pricing of $5 per million input tokens and $30 per million output tokens, plus subscription tiers.

This page covers how Fugu Ultra can assist with legal research and synthesis, the non-negotiable step of verifying every citation, and the confidentiality issues that come from routing your prompts to third-party vendors.

How Fugu Ultra fits research and synthesis

As a general assistant, Fugu Ultra can help at the exploratory and synthesis stages of research - the parts where you are organizing your own thinking rather than establishing what the law is. It can turn a messy question into a structured outline, summarize material you provide, and suggest angles to investigate.

What it cannot do is serve as an authority for what a case holds or whether a case exists. It is not connected to an authoritative, verified legal database by default, and it can generate plausible-looking citations that are wrong. Use it to accelerate your own work, then confirm everything in a primary source or a trusted legal research service.

• Draft research outlines and issue lists from a plain-English question.
• Summarize documents or opinions you supply, for faster review.
• Suggest counterarguments or lines of inquiry to pursue.
• Rephrase dense passages into clearer working notes.

Mandatory citation verification and the Mata lesson

The most important rule for AI-assisted legal research is that you verify every citation before you rely on it. AI systems, including orchestrators that route across multiple models, can fabricate case names, reporter cites, quotations, and holdings that look entirely real.

In Mata v. Avianca, a federal court sanctioned attorneys who submitted a brief citing cases that an AI tool had invented. The lesson is direct: a confident citation from any AI system is not evidence that the case exists or says what the tool claims. You must pull and read the actual source.

Because Fugu Ultra combines output from several models, a hallucinated citation from any single model in the chain can appear in the final answer. Orchestration does not eliminate this risk. Independent verification against the primary source is required every time.

• Pull and read every cited case, statute, or rule before relying on it.
• Confirm quotations word-for-word against the primary source.
• Never file a brief containing an unverified AI citation (see Mata v. Avianca).
• Assume any citation could be fabricated until you have checked it.

Confidentiality: your research prompts leave Sakana

Research prompts often contain sensitive facts - client names, matter details, or the theory of a case. With Fugu Ultra, those prompts are routed to third-party frontier vendors (Anthropic, OpenAI, or Google), so confidentiality and privilege depend on both Sakana's data-handling terms and every downstream vendor's terms.

Sakana has not stated that it offers a BAA or any specific certification for Fugu. Before entering client-identifying research questions, confirm in writing how prompts are stored, logged, and routed by every model that may process them, and whether any provider retains or trains on the data.

Under ABA Model Rule 1.6, you must make reasonable efforts to protect client information. A practical approach is to strip identifying details from research prompts, or to avoid the tool for privileged research until the multi-vendor data terms are confirmed.

Limitations and human review

Fugu Ultra can be confidently wrong on substance as well as citations. It may misstate a holding, miss a controlling authority, or blend jurisdictions without flagging the difference. It also has a knowledge cutoff and no reliable awareness of very recent decisions or rule changes.

A qualified attorney must review and take responsibility for any research output before it informs advice or a filing. Under ABA Model Rules 5.1 and 5.3, supervising lawyers are responsible for work produced with the help of technology and non-lawyer assistance.

None of this is legal advice, and Layer3 Labs is not a law firm. Treat AI research output as a draft to be checked, never as the final word.

• Check for missed controlling authority and mixed jurisdictions.
• Account for the model's knowledge cutoff on recent law.
• A responsible attorney must own the final research product.

Honest fit for legal research

Fugu Ultra is a reasonable fit for early-stage brainstorming, outlining, and summarizing documents you already have, when inputs are de-identified and every legal claim is verified. It is a poor fit as a primary legal research engine, because it is not a verified legal database and can fabricate authority.

It is also a poor fit when a firm needs a signed BAA or specific certification, which Sakana has not stated for Fugu, or when research prompts must contain client-identifying data that cannot be routed to third-party vendors. For authoritative citation retrieval, a dedicated legal research service remains the appropriate tool.

• Good fit: de-identified brainstorming, outlining, and document summarizing.
• Poor fit: authoritative citation retrieval or verifying that a case exists.
• Poor fit: privileged prompts before written routing/retention confirmation.