Which AI Model for Legal Work in 2026

For most legal work in 2026, the default is Claude Opus 4.8 — because legal work is citation-bound, and Opus is the only model whose behavior at depth matches the job's standard of proof. It holds a full contract bundle in its 1M-token window, retrieves the clause buried at page 400 more reliably than any flagship, and returns it quoted verbatim rather than paraphrased — pushing back when the text doesn't support a claim instead of inventing agreement. Switch to Gemini 3.1 Pro when discovery-scale volume makes economics decisive, GPT-5.5 when the matter turns on computed numbers, and Mistral when EU data residency governs. And treat all of it as leverage under attorney supervision — nothing below replaces legal judgment or review.

The reason legal work needs its own entry in this series is that it inverts the usual quality question. In most domains, a plausible paraphrase is acceptable output. In legal work, the paraphrase is the failure: the next reviewer needs the operative language itself, tied to where it appears, because meaning turns on exact words. That single requirement — verbatim grounding at depth — reorders the model rankings from almost every other task in the which AI model should you use hub.

It also raises the stakes on the failure mode every lawyer has now heard about: fabricated authority. Models asked for case law from memory produce confident citations to cases that don't exist. The mitigation is architectural, not hopeful — ground the model in real sources, demand quotes, verify every cite — and it's the frame for everything below.

How We Evaluated

This is a working matrix for legal workloads, not a leaderboard. The dimensions are the ones that predict whether a model's output survives the next reviewer — a partner, opposing counsel, a judge.

• Citation & quote discipline — whether findings arrive as verbatim quotes tied to their location, and whether the model resists asserting what the source doesn't support. The load-bearing dimension for legal work.
• Deep retrieval over long documents — how reliably the model finds the relevant clause when it sits at token 700,000 of a contract stack.
• Legal drafting quality — the prose standard of memos, letters, and clause language: precise, appropriately hedged, in professional register.
• Quantitative analysis — damages models, payment schedules, interest and exposure calculations, where predicted arithmetic is a liability and executed arithmetic is a requirement.
• Discovery-scale economics — cost behavior when the workload is thousands of documents, not one contract.
• EU data-residency story — where the data goes, and how short the conversation with the client's DPO is.

Honesty disclaimer. Public legal-AI benchmarks exist and move every release cycle; we deliberately don't quote percentages. The capability columns are qualitative buckets — Best-in-class, Strong, Good, Adequate — based on how these models behave on real contract stacks and research memos, consistent with their profiles across long-context document analysis. The only benchmark that matters is your own matter types, verified by your own attorneys.

The Decision Matrix

The story: Opus 4.8 leads the three dimensions that define legal work — quote discipline, deep retrieval, and drafting — which is why it's the recommended winner. Gemini 3.1 Pro matches the window at a mid cost tier, making it the volume pick, with softer citation behavior as the trade. GPT-5.5 owns the quantitative lane via its execution sandbox. Mistral trades capability ceiling for the EU residency story no US flagship offers.

Claude Opus 4.8: When It's the Right Call

Opus 4.8 is the default because it's built to the evidentiary standard legal work actually applies. Load a full acquisition document set into its 1M-token window and ask for every change-of-control provision, and it returns them quoted — the operative language, with its location — rather than summarized. Ask it to confirm a reading the text doesn't support, and it pushes back rather than agreeing. That pairing of best-in-class recall at depth with best-in-class quoting is precisely what citation-bound review rewards, and its drafting completes the loop: memos and clause language in a register that needs the least partner-level rewriting of any model.

Pick Claude Opus 4.8 when:

• The deliverable is quoted findings — contract review, diligence summaries, clause extraction, compliance mapping.
• The source material is huge: full contract bundles, records, and briefs held in one window without chunking.
• Drafting quality is load-bearing — the memo, the letter, the markup rationale.
• You're synthesizing provided sources into a research memo where every claim must trace to its authority.

Avoid Claude Opus 4.8 when:

• The matter turns on computed numbers — pair it with GPT-5.5's sandbox rather than asking Opus to be the calculator.
• The workload is discovery-scale and cost per document dominates — route the bulk pass to Gemini and reserve Opus for what it flags.

GPT-5.5: When It's the Right Call

GPT-5.5 earns its lane wherever legal work becomes quantitative. Damages models, prejudgment interest, payment waterfalls, royalty calculations, exposure scenarios — its execution sandbox runs real code against the actual figures and returns verified numbers instead of predicted ones, with the code available for audit. That's the difference between a settlement analysis you can defend and one you have to re-derive. It's also a strong reviewer and drafter within its 400K-token window, and its high reasoning effort helps on intricate multi-factor analysis.

Pick GPT-5.5 when:

• The matter involves computed exposure: damages, interest, schedules, valuation scenarios.
• You want an auditable calculation trail — the code that produced each figure.
• The reasoning task is deep but the documents fit comfortably under 400K tokens.

Avoid GPT-5.5 when:

• The document set exceeds its window — that's the 1M-token models' lane.
• The deliverable is quoted findings from a giant record, where Opus's quote discipline is the stronger guarantee.

Gemini 3.1 Pro: When It's the Right Call

Gemini 3.1 Pro is the volume play. It matches Opus's 1M-token window at a mid cost tier, which changes the economics of discovery-scale work: first-pass relevance sweeps, categorizing thousands of documents, surfacing candidate hot documents, summarizing depositions at scale. Its native multimodality also earns real work when the record includes exhibits that aren't clean text — scanned agreements, charts, photographs of documents. The trade-off is the one that matters most in this domain: it paraphrases more than it quotes, so treat its output as triage that routes documents to a stricter second pass, not as citable findings.

Pick Gemini 3.1 Pro when:

• The workload is discovery-scale and cost per document decides feasibility.
• The record is mixed-media — scans, exhibits, charts — that needs native visual reading.
• The deliverable is triage and synthesis, with a stricter model or a human doing citation-grade follow-up.

Avoid Gemini 3.1 Pro when:

• Output goes into a filing or client deliverable without a verification pass — its paraphrasing is a liability at that standard.

Mistral Large 3: When It's the Right Call

Mistral's case is jurisdictional. Its hosted services run in the EU, which for European firms and EU-client matters converts the data-residency analysis from a negotiation into a checkbox — and its European-language output holds legal register in French, German, Spanish, and Italian in a way US-trained models often miss. Capability-wise it's a competent generalist rather than a retrieval specialist, so its role in most legal stacks is scoped: client-confidential EU matters, multilingual client communication, and structured extraction where its reliable JSON keeps pipelines clean.

Pick Mistral when:

• EU data residency is a client or regulatory requirement.
• The work is multilingual across European languages and register matters.
• Stricter still: self-host an open model inside firm infrastructure — the private and self-hosted matrix covers that path.

Avoid Mistral when:

• The task is deep retrieval over a giant record — the 1M-token models lead decisively.

Which to Pick by Sub-Segment

Contract review and due diligence

Opus 4.8, and it's the clearest call in the matrix. Full bundle in one window, findings quoted with locations, playbook deviations flagged against the actual language. For M&A-scale volume, run Gemini 3.1 Pro as the first-pass sweep and route flagged documents to Opus for citation-grade extraction.

Legal research memos

Opus 4.8 for synthesis over sources you provide. For finding current authority, ground the model in a legal research database or retrieval layer first — no general model's memory is citable — then have Opus write the memo from what was retrieved. Every cite gets human verification; the anti-hallucination workflow is not optional in this sub-segment, it's the sub-segment.

Drafting: memos, letters, clause language

Opus 4.8 for register and precision, with GPT-5.5 a strong second. Draft against a playbook or precedent clause pasted into context rather than from the model's memory of "standard" language — grounded drafting inherits your firm's positions instead of the internet's averages.

Damages, exposure, and settlement analysis

GPT-5.5, for the sandbox. Every figure computed by executed code, every assumption explicit, the calculation auditable. Hand the verified numbers to Opus when the analysis needs to become a narrative a mediator or client will read.

Discovery and document triage

Gemini 3.1 Pro for economics and multimodal reading at scale, feeding a stricter second pass. At extreme volume with strict confidentiality, a self-hosted open model inside firm infrastructure is the alternative — the open-weight guide covers the trade-offs.

EU and cross-border matters

Mistral hosted in the EU, or self-hosted open weights where nothing may leave the firm. Reserve flagship calls for anonymized or public-material work.

Sample Prompt for the Recommended Winner

A copy-paste prompt for Claude Opus 4.8 contract review. The structure enforces the discipline that makes AI legal work safe: quotes only, locations always, gaps flagged rather than filled.

You are assisting a licensed attorney with a first-pass contract review.
I am providing the full text of the agreement below. Work ONLY from the
provided text — do not rely on your memory of "standard" contract terms.

Tasks:
1. Extract every provision addressing: (a) termination, (b) indemnification,
   (c) limitation of liability, (d) change of control, (e) governing law.
2. For each provision found, return: the section number, the VERBATIM quoted
   language, and a one-sentence plain-English description. Do not paraphrase
   the operative language — quote it exactly.
3. Flag any provision that deviates from the checklist positions I've pasted
   below the agreement, quoting both the checklist position and the
   deviating language side by side.
4. If a category has no provision in the agreement, state "No provision
   found" — do not infer or suggest what the agreement "likely intends."

Rules:
- Every claim must be supported by a verbatim quote from the provided text.
- If language is ambiguous, say so and present the competing readings —
  do not resolve ambiguity by assumption.
- This is a first-pass review for attorney evaluation, not legal advice.

[PASTE AGREEMENT TEXT]
[PASTE CHECKLIST POSITIONS]

Two design choices do the work. The verbatim-quote requirement converts the model's output into checkable evidence — a quote either appears in the document or it doesn't, which makes verification a ten-second search instead of a re-read. And the "no provision found — do not infer" rule closes the gap-filling failure mode, where a helpful model drafts the clause it expected to find. That's the same grounding discipline from our hallucination-prevention guide, tuned to the domain where its absence is sanctionable.

Closing

Legal AI in 2026 is a routing problem with an unusually strict acceptance test: does the output arrive quoted, located, and verifiable? For most matters that makes Claude Opus 4.8 the default — the model whose depth retrieval and citation discipline were effectively built to this standard. Route around it deliberately: Gemini 3.1 Pro when discovery volume makes mid-tier economics decisive, GPT-5.5 when the matter turns on executed math, Mistral when EU residency governs, and self-hosted open weights when nothing leaves the firm.

Then hold the line on process: ground every task in real sources, require quotes, verify every citation, and keep attorney judgment where it's always been. For the cross-task framework behind these picks, see the AI model selection guide and the which AI model should you use hub — and for ready-made structures, the prompt generator for lawyers builds review, research, and drafting prompts with the grounding rules already in place.