Glossary
Reasoning Model — Definition for Lawyers
A reasoning model works through intermediate steps before it answers. What adaptive thinking and effort settings mean for legal analysis, cost, speed and hidden reasoning.
On this page
A reasoning model is a large language model that spends extra computation working through a problem before it answers, generating and checking intermediate steps rather than predicting the reply in one pass. As of September 2026 every frontier model reasons adaptively and lets you set how hard it tries; the trade is slower, dearer answers for fewer errors on multi-step legal analysis, not zero errors.
How reasoning models differ from ordinary chat models
Earlier chat models answered in a single forward pass: each token predicted from the last, with no chance to backtrack. A reasoning model first writes a private working draft, sometimes thousands of tokens long, in which it plans, tries alternatives, spots contradictions and revises, and only then produces the answer you see. The large language model entry covers the mechanics underneath; what matters here is that the working draft costs tokens and time, and that vendors now treat it as a dial rather than a switch.
Anthropic's current line makes thinking "adaptive", controlled by an effort parameter with five settings from low to max; on Claude Fable 5.1 thinking is always on and the raw chain of thought is never returned. OpenAI's GPT-6 Astra, public to paid users since 4 September 2026 and covered on the ChatGPT for lawyers page, uses a "recurrent depth" technique that obscures some or all of the model's reasoning. Claude Opus 5, launched 24 July 2026, goes a step further and, in Anthropic's words, "verifies its own work without being told to". That has a practical consequence: prompts written in 2025 that end with "now add a verification step" cause over-verification on Opus 5, and longer, slower answers.
What the effort dial buys, and what it costs
| Effort setting | Typical legal task | What you gain | What you pay |
|---|---|---|---|
| Low | Summarise a 20-page lease; extract dates and parties; reformat a clause | Speed and price; answers in seconds | Little; these tasks rarely need multi-step reasoning |
| Medium | First-pass NDA redline against a playbook; issue-spot a set of board minutes | Fewer missed cross-references | More output tokens and a longer wait |
| High and above | Tax structuring across three jurisdictions; reconciling conflicting precedent; statutory interpretation with nested conditions | Materially fewer logical slips where steps depend on each other | Minutes rather than seconds, and a bill to match |
The benchmark that priced this trade is HAQQ's June 2026 test of 3,000 answers across 300 legal tasks. GPT-5.5 had the lowest hallucination rate of any model at 3%, and was also the slowest at 134 seconds a task and the dearest at $0.082; the cost spread across the models tested was 90x. OpenAI's older o3 reasoning model hallucinated 32% of the time on the same test, so "reasoning" on the label guarantees nothing. Across all 3,000 answers, 24% cited or misapplied law that did not support the claim. Reasoning reduces the hallucination rate; it does not remove the verification step, and the benchmarks guide explains why headline numbers differ so much between tests.
The hidden-reasoning problem for lawyers
A reasoning model's working draft looks like an audit trail, and it is tempting to file it as one. Do not. On Fable 5.1 you never see it; on GPT-6 Astra part of it is not expressed in text at all; and on models that show a "thinking" summary, the summary is generated after the fact and can be as confident, and as wrong, as the answer. Safety researchers worry about monitorability. You should worry about explainability. If a client, a court or a regulator asks why the advice says what it says, the reasoning has to be yours, reconstructed from sources you have read, not a paste of the model's scratchpad.
The workaround is to ask for the reasoning inside the answer, as structure you can check:
Before your conclusion, set out:
1. The facts you rely on, quoted from the documents provided.
2. Each legal rule you apply, with the section or case it comes from.
3. Any point where two readings are possible, and why you prefer one.
4. What you could not verify from the materials.
Then give the conclusion in no more than 150 words.
That prompt works at any effort level and on any vendor's model, and it turns hidden reasoning into checkable reasoning. The prompt engineering guide has the longer version with worked examples.
When to turn it up
Use high effort for anything with dependencies: a multi-jurisdiction question, a contract where clause 14 changes the meaning of clause 3, a chronology with contested dates. Leave it low for extraction, translation, summarising and formatting, where you would be paying for thinking the task does not need. Set the effort per task rather than per firm, and record which setting produced a piece of work that goes out under your name. Module 1 of the AI Lab for Lawyers course, on fundamentals and prompting, is where these settings get worked through on Claude, ChatGPT and Gemini side by side.
Frequently asked questions
What is a reasoning model in AI?
A reasoning model is a large language model that spends extra computation before answering, generating and checking intermediate steps rather than predicting the reply in one pass. As of September 2026 the frontier models from Anthropic, OpenAI and Google all reason adaptively, deciding how much to think based on the task, and Anthropic lets you set an effort level from low to max. The cost is time and tokens; the benefit is fewer logical slips on problems with several dependent steps.
Do reasoning models hallucinate less in legal research?
Less, not never. In HAQQ's June 2026 benchmark of 3,000 answers, GPT-5.5 had the lowest hallucination rate at 3% and was also the slowest and most expensive model tested, while OpenAI's older o3 reasoning model hallucinated 32% of the time. Across every model, 24% of answers cited or misapplied law that did not support the claim. Reasoning changes the odds; it does not change the duty to read every citation before it goes out under your name.
Should lawyers always use the highest thinking or effort setting?
No. High effort pays on tasks with dependencies, such as a contract where one clause changes the meaning of another, a tax question across three jurisdictions or a contested chronology. For extraction, summarising, translation and reformatting it adds minutes and cost without improving the result. Choose the setting per task, and note which setting produced any piece of work that leaves the firm.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Guides
How Large Language Models Work, Explained for Lawyers
How LLMs work, explained for lawyers: tokens, prediction, training versus inference, context windows, knowledge cutoffs, reasoning modes and the September 2026 model line-up.
8 min read
Glossary
Hallucination (AI) — Definition for Lawyers
An AI hallucination is fluent, confident, false output: an invented case, a misread holding or a fake quote. 2,041 court decisions now involve one. What to know.
5 min read
Glossary
Large Language Model (LLM) — Definition for Lawyers
A large language model is a neural network trained to predict the next token, and it powers every legal AI tool. How LLMs are built, the models of September 2026 and what follows for reliability and confidentiality.
4 min read
Guides
Prompt Engineering for Lawyers: Techniques That Actually Work
Prompt engineering for lawyers as a method: six elements, six complete legal prompts (NDA review, research memo, clause redraft, client email) and iteration.
8 min read
Questions
Which AI Has the Lowest Hallucination Rate for Legal Work?
The lowest hallucination rate AI for legal work as of September 2026: GPT-5.5 at 3% on HAQQ, Claude Opus 4.8 best overall, and why no model reaches the zero courts demand.
6 min read
Glossary
Context Window — Definition for Lawyers
A context window is the text an AI model can hold at once. As of September 2026 it is 1 million tokens on Claude, GPT-6 and Gemini, about 555,000 words.
4 min read