Reasoning Models for Legal Work: New Prompting Rules (and Why They Can Hallucinate More)

Somewhere in your firm’s prompt library is a line that reads “Think step by step, then double-check your answer before responding.” In 2023 that line measurably improved GPT-4: Thomson Reuters’ CoCounsel team showed “Before you answer, think through your reasoning step-by-step” turning a wrong LSAT-style answer into the right one. In 2026 OpenAI says the same instruction is “unnecessary” for its reasoning models, and Anthropic says the second half should be deleted from any prompt for Claude Opus 5 because such instructions “cause over-verification”.

Reasoning models for legal work are a different animal from the chat models the frameworks were written for, and most legal prompting guides have not caught up. Clio, the Singapore Academy of Law and Thomson Reuters still teach chain-of-thought; OpenAI and Anthropic now tell you to stop. Both camps are right, about different models. And the data nobody puts on the slide: the models that reason best are not the models that hallucinate least.

What a reasoning model does that a chat model does not

A chat model answers in one pass. A reasoning model first generates a hidden or partly visible chain of “thinking” tokens, plans, tries alternatives, and only then writes the answer. Those thinking tokens count against the context window and are billed as output, so a reasoning model is slower and dearer.

OpenAI’s guide draws the division of labour: o-series models are “the planners”, GPT models “the workhorses”. Reasoning models suit “Navigating ambiguous tasks” and “reasoning over complex documents that have hundreds of pages of dense, unstructured information—things like legal contracts, financial statements, and insurance claims” (OpenAI reasoning best practices). GPT-5.1 exposes the dial: its none reasoning setting behaves “much more similar in usage to GPT-4.1”, so one product can be either kind depending on a setting.

OpenAI: “think step by step” is unnecessary for o-series models

The verbatim instruction: “Avoid chain-of-thought prompts: Since these models perform reasoning internally, prompting them to ‘think step by step’ or ‘explain your reasoning’ is unnecessary.” The rest of the list is short: “Keep prompts simple and direct”; “Use delimiters for clarity”; “Try zero shot first”; “Provide specific guidelines”; “Be very specific about your end goal.”

The legal examples OpenAI cites are the interesting part. Hebbia reported that “o1 enabled Matrix to easily identify baskets available under the restricted payments capacity in a credit agreement, with a basic prompt”: a task that defeats keyword search because baskets are defined by cross-reference and arithmetic, not by a heading. Blue J, the tax research platform, reported a “4x improvement in end-to-end performance” from GPT-4o to o1. In both cases the prompt got simpler and the model did the planning.

Anthropic: Opus 5 “verifies its own work without being told to”

Claude Opus 5 “verifies its own work without being told to. If your prompt contains explicit verification instructions… remove them: instructions like these cause over-verification.” The phrases to delete are “double-check your answer” and “re-verify before responding” (Anthropic, prompting Claude Opus 5).

Two side effects matter. Opus 5’s default responses “run longer than prior Opus models’”, so ask for brevity. And it can over-read scope, so Anthropic publishes a scope-control instruction that reads like a partner’s note to a keen associate.

Keep responses focused, brief, and concise. Keep disclaimers and caveats short, and spend most of the response on the main answer. When asked to explain something, give a high-level summary unless an in-depth explanation is specifically requested.

Deliver what was asked, at the scope intended. Make routine judgment calls yourself, and check in only when different readings of the request would lead to materially different work. If the request seems mistaken or a better approach exists, say so in a sentence and continue with the task as asked rather than quietly narrowing, widening, or transforming it. Finish the whole task, and stop short of actions that are clearly beyond what was asked.

Anthropic’s general best-practices page confirms the exception. For most models it still says “Append something like ‘Before you finish, verify your answer against [test criteria].’ This catches errors reliably”, and then: “Claude Opus 5 is the exception.” One wrinkle for the previous generation: with extended thinking off, Opus 4.5 “is particularly sensitive to the word ‘think’”, so use “consider”, “evaluate” or “reason through” (Anthropic prompting best practices).

When chain-of-thought still helps

The frameworks are not wrong; they are dated. Thomson Reuters’ 2023 tests on GPT-4 showed “Before you answer, think through your reasoning step-by-step” fixing a wrong answer and “Double-check your answer and fix any problems you find” correcting another (Thomson Reuters). The SAL/Microsoft guide still advises that “For more complex tasks, including the phrase ‘think through this step-by-step’ can improve the output”, and Clio notes a benefit that survives: visible reasoning “makes verification easier.”

So the rule is model-dependent:

• On a non-reasoning model or mode (GPT-4.1, GPT-5.1 with reasoning none, Claude with thinking off, most Copilot surfaces), chain-of-thought and a self-check line still earn their place.
• On a reasoning model, replace the procedure with a goal and a success criterion. Anthropic’s middle-ground advice: “Prefer general instructions over prescriptive steps. A prompt like ‘think thoroughly’ often produces better reasoning than a hand-written step-by-step plan.”

The catch: reasoning models “make more claims overall” and hallucinate more

Here is the number for every slide about reasoning models. On OpenAI’s own PersonQA benchmark, o3 hallucinated on 33% of questions and o4-mini on 48%, against 16% for o1. OpenAI’s explanation was that the newer models “make more claims overall” and so produce “more accurate claims as well as more inaccurate/hallucinated claims” (TechCrunch, April 2025).

The legal evidence points the same way. The one published randomised controlled trial of a reasoning model in legal work that I know of (Schwarcz and colleagues, 2025) gave upper-level law students Vincent AI (a retrieval-grounded tool), OpenAI’s o1-preview, or no AI, across six tasks. Both tools “significantly enhance legal work quality”, with productivity gains of 50% to 130% on five of six tasks. The hallucination count: Vincent AI 3, students without AI 4, o1-preview 11 (Michigan Law repository). Better work, and nearly three times the fabrications of a student with a library card. Damien Charlotin, who maintains the hallucination-cases database, is blunt: “I don’t really buy the advances in terms of reduced hallucinations for newer models.”

What the legal benchmarks show

The gains are real, which is why the hallucination figures deserve attention rather than dismissal.

LegalOn’s figures come from 494 review decisions across five contract types with no customisation (Artificial Lawyer, March 2026). Linklaters’ verdict on its own results is the one to keep: the tools “should not be used for English law legal advice without expert human supervision” (Linklaters). Legal AI benchmarks explained sets out what each measures.

Deep Research modes: what they are for and how to verify them

Deep Research gives the model a browser and a time budget. OpenAI launched it on 2 February 2025; a report “may take anywhere from 5 to 30 minutes” and arrives with a source list (OpenAI).

Use it for the first hour on an unfamiliar regulatory topic: who regulates, which instruments, what timeline, who disagrees. Do not use it for the citation you will file. Harvey’s diligence team, which benchmarks these agents: “Their bias is to search efficiently, not completely. Diligence requires reversing this intuition.” And dates go stale: the EU AI Act’s Annex III deadline moved to 2 December 2027 in July 2026, so a report built on earlier sources is wrong. For authority you will rely on, legal research without hallucinations uses grounded platforms with a citator.

Produce a briefing for a lawyer new to [EU AI Act obligations for deployers] as at [today's date]. Cover: the governing instruments with official links; who is regulated and who enforces; the compliance timeline with exact dates and any recent deferrals; the three most-cited practitioner summaries, linked; the open questions; a glossary of ten terms.
Prioritise primary sources (official journals, regulator pages) over news. Put the source next to every date and threshold. Finish with a section headed "What I did not find". No more than 1,500 words plus the source list.

A decision table: which model, which prompt style, which task

ChatGPT vs Claude for lawyers covers the brand question; one data point is that in Chroma’s long-context tests Claude models “tend to abstain when uncertain” while GPT models showed “the highest rates of hallucination” with distracting passages.

Prompt templates for reasoning models: brief, constraints, success criterion

The old prompt was a procedure. The new one is a brief for a capable senior associate: the deliverable, what it must not contain, and how you will judge it.

Goal: a complete list of every provision in this MSA that allocates risk away from my client, the customer, so nothing is missed at tomorrow's negotiation call.
Success criterion: completeness matters more than brevity. I would rather see twenty candidate issues with three false positives than ten issues with one omission.
For each provision: quote the operative words, give the clause reference, and say in one line why it shifts risk. Finish with the three provisions you would raise first and why.
Constraints: [English] law; use only the document; where a clause depends on a schedule not attached, say so.

<document>[paste]</document>

Add a footer asking for CONFIDENCE, WHAT I DID NOT HAVE and WHAT TO VERIFY FIRST; models rarely say “Low”, so treat “Medium” as “Low”. It is the fastest way to turn fluency into a to-do list.

Where to go next: the full framework is in prompt engineering for lawyers, the twelve failures that end up in sanctions orders are in legal prompting mistakes, and the templates live in the prompt library; the prompting hub lists the rest. AI Lab for Lawyers is taught on current models, so every habit you learn there matches the tools you will be using next month.