Reviewed by Jonathan West · Updated Jul 4, 2026

Fugu Ultra for Paralegals

How Fugu Ultra can support paralegal tasks, why supervising-attorney review is required, and what its routing means for client confidentiality.

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 one OpenAI-compatible API and returns a single combined answer. Its coordination approach builds on two ICLR 2026 papers, TRINITY and Conductor.

Fugu Ultra is the higher-performance tier of the Fugu family, aimed at harder tasks, while standard Fugu prioritizes lower latency and cost. It 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, with subscription tiers.

This page covers how Fugu Ultra can support common paralegal work - document review, discovery summaries, deposition prep, and cite-checking - always under supervising-attorney review, plus the confidentiality issues that come from routing prompts to third-party vendors.


Paralegal tasks Fugu Ultra can support

As a general assistant, Fugu Ultra can speed up the organizing and summarizing work that fills a paralegal's day. It works from the materials you provide and produces drafts for a supervising attorney to review.

The tasks below can save time, provided a lawyer reviews the output and every factual and legal detail is verified against the source.

  • First-pass document review - grouping, tagging, and flagging documents for a lawyer.
  • Drafting discovery summaries from materials you supply.
  • Building deposition prep outlines and question lists to hand to an attorney.
  • Cite-checking support - listing citations to verify, not confirming them.
  • Summarizing long records into short briefings for the team.
Fugu Ultra helps organize and draft. A supervising attorney reviews the output; the tool does not replace legal judgment.

Setting up AI-assisted paralegal work with supervision and confidentiality intact? We will help.

Book a Consultation

Supervising-attorney review is required

Paralegal work supports a lawyer, and AI-assisted paralegal work is no different. Under ABA Model Rule 5.3, lawyers are responsible for the conduct of non-lawyer assistants, and that responsibility extends to work produced with AI tools. A paralegal using Fugu Ultra should treat every output as a draft for the supervising attorney to review and approve.

This matters most for anything that resembles legal judgment - what a case holds, whether a document is responsive, how to frame a deposition question. The tool can produce a confident answer that is wrong, and only a lawyer can make the final call.

A clear internal rule - that AI output is reviewed by the supervising attorney before it is relied on or shared outside the team - keeps the supervision structure intact.

  • The supervising attorney reviews and approves AI-assisted output (ABA Rule 5.3).
  • AI output is a draft, never a final legal determination.
  • Anything resembling legal judgment goes to the lawyer for the final call.

Confidentiality: case materials leave to third-party vendors

Paralegals handle some of the most sensitive material in a matter - discovery documents, client records, and case files. With Fugu Ultra, anything you submit is routed to third-party frontier vendors (Anthropic, OpenAI, or Google), so confidentiality and privilege depend on both Sakana's 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 case material, confirm with the supervising attorney and in writing how the data is stored, logged, and routed by every model in the chain, and whether any provider retains or trains on it.

Under ABA Model Rule 1.6, the firm must make reasonable efforts to protect client information, and that duty flows to non-lawyer staff. Practical approaches include working from de-identified material, or not using the tool for privileged case documents until the multi-vendor terms are confirmed. Also check whether a protective order or discovery agreement limits sending materials to third-party processors.

Case materials submitted to Fugu Ultra leave Sakana and reach third-party models. Sakana has not stated a BAA or certification for Fugu - confirm routing and retention in writing, and check protective orders, before entering client-identifying material.

Accuracy limits and cite-checking

Fugu Ultra can be confidently wrong. In document review it may misclassify a document; in a discovery summary it may misstate a fact; in cite-checking it may claim a citation is valid when it is not, or invent one entirely.

This is why cite-checking with AI must mean verifying against the primary source, not trusting the tool's word. Courts have sanctioned lawyers for filing AI-fabricated cases, as in Mata v. Avianca, and a cite-check that relies on the AI alone offers no protection. A paralegal should pull each source and confirm it independently.

Every summary, tag, and flag the tool produces should be treated as a hypothesis for a human to confirm, and the supervising attorney owns the result.

  • Verify every citation against the primary source - never trust the AI's confirmation.
  • Confirm document classifications and summarized facts against the originals.
  • Treat every AI output as a hypothesis for human confirmation.
Cite-checking with AI means verifying each source yourself. Courts have sanctioned lawyers over AI-fabricated cases (Mata v. Avianca).

Honest fit for paralegal work

Fugu Ultra is a reasonable fit for organizing, summarizing, and drafting support tasks where materials can be de-identified and a supervising attorney reviews the output. It can save real time on first-pass review and outline-building.

It is a poor fit for authoritative cite-checking - it cannot confirm a citation exists - and for handling privileged case materials before the multi-vendor routing and retention terms are confirmed in writing. It is also a poor fit where a firm needs a signed BAA or specific certification that Sakana has not stated for Fugu, or where a protective order restricts third-party processing. For high-volume, low-complexity summarizing, the lower-cost standard Fugu tier may be more economical than Ultra.

  • Good fit: de-identified organizing, summarizing, and draft support with attorney review.
  • Poor fit: authoritative cite-checking or confirming a case exists.
  • Poor fit: privileged materials before written routing/retention confirmation.
  • Poor fit: matters under a protective order barring third-party processing.

Frequently Asked Questions

  • It can support first-pass document review, discovery summaries, deposition prep outlines, and cite-checking support, all from materials you supply. Every output is a draft for a supervising attorney to review.
  • No. The tool can claim a citation is valid when it is not, or invent one. Cite-checking means verifying each source against the primary record yourself. Courts sanctioned lawyers over AI-fabricated cases in Mata v. Avianca.
  • Yes. Under ABA Model Rule 5.3, lawyers are responsible for non-lawyer assistance, including AI-assisted work. A supervising attorney should review and approve the output before it is relied on.
  • Not until the terms are confirmed. Materials are routed to third-party vendors, and Sakana has not stated a BAA or certification for Fugu. Confirm routing and retention in writing, check protective orders, and prefer de-identified material.
  • Standard Fugu prioritizes lower latency and cost, so for high-volume, low-complexity summarizing it may be more economical than Ultra, which lists $30 per million output tokens. Compare against your workload.

Support your paralegal team safely

Book a free 30-minute AI workflow audit with Layer3 Labs. We will help you set up an AI-assisted paralegal process that keeps supervision, cite-verification, and client confidentiality intact.

/ai-workflow-audit