AI Agents for Lawyers

Comparisons

Claude vs ChatGPT for Lawyers (2026): Which Should You Use?

Claude vs ChatGPT for lawyers in 2026, compared task by task: drafting, research, agents, data terms by plan, prices and benchmarks, plus a pick by firm type.

By Dr. Niklas SchmidtUpdated 13 November 202610 min read
On this page
  1. The short answer, task by task
  2. Benchmarks: the winner depends on who wrote the test
  3. Data terms by plan: the factor that should decide close calls
  4. What each costs
  5. The legal offerings: Claude for Legal vs Astra for Law
  6. Who is actually using which
  7. Recommendation by firm type
  8. A bake-off you can run in an hour
  9. What to do this week

As of September 2026, Claude is the better default for most lawyers: it holds the top two places on the Vals Index, produces stronger first-pass redlines and ships an open-source legal toolkit. ChatGPT wins where a firm needs GPT-6 Astra's computer-use agent, already runs ChatGPT Enterprise, or wants Astra for Law's US case-law index. Data terms, not benchmarks, should settle the close calls.

The short answer, task by task

The two products have converged on paper. Claude Fable 5.1 and GPT-6 Astra both cost $10 per million input tokens and $50 per million output tokens on the API, both take roughly a million tokens of context (about 555,000 words on Anthropic's tokeniser) and both return up to 128,000 tokens of output. The differences show up in what each does with a legal task.

TaskClaude (Fable 5.1 / Opus 5 / Sonnet 5)ChatGPT (GPT-6 Astra / GPT-5.6)Edge, as of Sep 2026
Contract redliningAnthropic says Opus 5 first-turn NDA redlines scored "nearly double Opus 4.8"; GC AI's RedlineBench rose to 57.0 with Fable 5.1GPT-5.6 Sol scored 92.7% on Harvey's BigLaw Bench, which a critic notes "does not actually test case law research"Claude, narrowly
US case-law researchOpus 5 and Fable 5.1 sit second and third on the Vals Legal Research Bench (15 Sep 2026)Astra for Law's Legal Search Index reportedly lifted Vals research accuracy to 54% from 38.7% for GPT-6 with plain web searchChatGPT, if you buy Astra for Law and practise US law
Long documents (data rooms, transcripts)1M context on all three current models~1.05M context on AstraTie
Citation hallucination rateOpus 4.8 first overall on HAQQ (30.02/35) with ~12% citation hallucinationGPT-5.5 lowest citation hallucination on HAQQ (3%), but slowest (134 s) and most expensive ($0.082 per task)ChatGPT on that test; both need verification
Long-horizon legal agentsFable 5 13.3% and Opus 5 11.7% on Harvey's strict LAB all-pass metric (18 Sep 2026)GPT-5.5 2.1% on the same metricClaude
Desktop and file workCowork merged into the Claude app on 17 Sep 2026; Claude Docs and Slides in betaChatGPT Work (9 Jul 2026) runs "independently for hours" with 1,400+ plugins and approval gatesTie
Browser and computer useClaude for Chrome (GA 18 Dec 2025), prompt-injection success cut from 23.6% to 11.2%GPT-6 Astra is the computer-use model; off by default in Enterprise until an admin enables itChatGPT, once IT approves
WordFirst-party Claude add-in with tracked changes, reported as included in paid plans; Microsoft's Word Legal Agent also runs on Anthropic modelsNo first-party legal Word add-in as of September 2026; Copilot is the Microsoft routeClaude

The Claude review and the ChatGPT review go through each product's features in depth. This page is about choosing between them.

Benchmarks: the winner depends on who wrote the test

Four leaderboards, four different answers. On the Vals Index refreshed on 15 September 2026 across 58 models, Claude Fable 5.1 is first, Claude Opus 5 second and GPT-6 Astra third. On the academic LegalBench (16 September 2026), Fable 5 scores 88.56% and Fable 5.1 88.51%, with Gemini 3.1 Pro at 87.40%. On Harvey's BigLaw Bench, OpenAI models occupy the top three places (GPT-5.6 Sol 92.7%, GPT-5.5 91.7%, GPT-5.4 91.0%) ahead of Claude Opus 4.7 at 90.9%. On the vendor-run GC AI In-House Legal Bench (May 2026), GPT-5.5 scored 79.8% against 68.4% for Opus 4.7.

Then there is the HAQQ benchmark of 5 June 2026: 3,000 graded answers across 300 tasks, 51 practice areas and 20-plus jurisdictions. Claude Opus 4.8 won 130 of 300 tasks and finished first; GPT-5.5 had the lowest citation-hallucination rate at 3% but took 134 seconds and $0.082 per task. Across all models, 24% of answers "cited or misapplied law that didn't support claims".

Two lessons for a buyer. First, the vendor that publishes a benchmark tends to win it, so read how legal AI benchmarks work before quoting one to a partner meeting. Second, the gap between Claude and ChatGPT on any single test is smaller than the gap between either model and a lawyer who checks the output. Harvey's own LAB numbers make the point: the best model in the world completes only 13.3% of long legal tasks without a single rubric failure.

Data terms by plan: the factor that should decide close calls

Model quality moves monthly. Retention terms move slowly and carry professional consequences. Here is the position as of September 2026, from Anthropic's retention documentation and OpenAI's enterprise privacy page.

TierTrains on your content?RetentionZero data retention available?
Claude Free, Pro, MaxYou must choose (since 28 Aug 2025)Opted in: 5 years; opted out: 30 daysNo
Claude Team, EnterpriseNeverFlagged content up to 2 years; Compliance API transcripts 6 years by defaultChat interface: no. Claude Code under Enterprise: yes
Claude API, Bedrock, VertexNeverNot retained by default, except Fable 5.1 and Mythos 5.1, which require 30-day retentionYes, per organisation, but not for Fable or Mythos without express authorisation
ChatGPT Free, Plus, ProYes, unless you opt out in Data ControlsStandardNo
ChatGPT Business, Enterprise, EduNo, by defaultStandard; Compliance API audit trailsContractual for eligible customers
OpenAI API and Astra for LawNo, by default30 days for abuse monitoringContractual; Astra for Law is offered with a zero-data-retention API

Two court rulings sharpen the point. In United States v. Heppner (S.D.N.Y., opinion 17 Feb 2026), Judge Rakoff held that a defendant's research memos written with consumer Claude were neither privileged nor work product, because "Claude is not an attorney" and the terms allow logging and disclosure. In the OpenAI copyright litigation, 20 million de-identified ChatGPT logs were ordered produced (affirmed 5 Jan 2026). Chat logs are discoverable ESI whichever vendor holds them; the privilege analysis is the same for both.

The practical rule: no client material on any consumer tier of either product. Switch it off in the settings using the step-by-step privacy guide, and buy a business tier before the first matter touches the tool. The confidentiality checklist covers the other vendors.

What each costs

PlanClaudeChatGPT
FreeSonnet 5 default; no Fable 5.1GPT-5.5 Instant
IndividualPro $20/month ($17 annual); Max $100 or $200Go ~$8 (regional); Plus $20; Pro $100 or $200
Team$25 standard or $125 premium per seat, 5-seat minimumBusiness $25 per seat monthly ($20 annual); Premium $125/$100
EnterpriseCustomCustom; GPT-6 Astra off until an admin enables it
API (in/out per million tokens)Fable 5.1 $10/$50; Opus 5 $5/$25; Sonnet 5 $2/$10; Haiku 4.5 $1/$5GPT-6 Astra $10/$50 (cached input $1)

For most practices the honest comparison is Claude Team at $25 against ChatGPT Business at $25: identical price, both non-training, both administrable. The legal AI pricing guide puts these numbers next to Harvey, Legora and CoCounsel.

Anthropic went first. Its legal plugin for Cowork launched on 3 February 2026 and knocked 18% off Thomson Reuters and 14% off RELX in a day. By 12 May 2026 it had grown into Claude for Legal: 12 practice-area plugins, 90-plus agents and skills, 20-plus MCP connectors (iManage, NetDocuments, DocuSign, Ironclad, Relativity, CoCounsel, LexisNexis, Harvey), all open source under Apache 2.0, with every output labelled "a draft for attorney review".

OpenAI answered on 17 September 2026 with Astra for Law: GPT-6 Astra configured for legal work, a Legal Search Index of 230 million-plus URLs of US law built with CourtListener and the Free Law Project, 26 partner plugins plus 47 community plugins, a zero-data-retention API and design partners Latham & Watkins, Sullivan & Cromwell, Ropes & Gray and Cooley. It is days old, and there are almost no independent evaluations yet. The head-to-head on the two legal offerings goes into the detail.

Who is actually using which

The ILTA 2026 Technology Survey of 500-plus firms (14 September 2026) puts Claude at 44% of firms and ChatGPT Enterprise at 22%, both behind Microsoft 365 Copilot at 76%. Freshfields rolled Claude out to 5,700 staff across 33 offices and saw roughly 500% usage growth in six weeks. Quinn Emanuel's Christopher Kercher described the breakthrough as "treating Claude like a member of the case team". On the other side, Gunderson Dettmer built ChatGD on OpenAI in 2023, and Latham, Sullivan & Cromwell, Ropes & Gray and Cooley are Astra for Law's design partners.

Harvey uses both: Claude Opus 5 went live in Harvey on 24 July 2026 and Fable 5.1 on 1 September, alongside GPT-5.6 Sol. So do Thomson Reuters (CoCounsel Legal is built on Anthropic's Claude Agent SDK) and LexisNexis (Anthropic, Google and OpenAI models). At platform level, the choice is increasingly made for you.

Neither vendor is immune to embarrassment. Where a hallucination decision names the tool, ChatGPT reportedly accounts for about half and Claude about 15%, a ratio that tracks usage more than quality. Anthropic's own counsel apologised to a federal court on 15 May 2025 after Claude produced a citation with the wrong title and authors.

Recommendation by firm type

Your situationPickWhy
Solo or small firm, Microsoft-based, no legal platformClaude Team (5-seat minimum) or Pro with training offStronger drafting; Claude for Legal plugins at no extra cost; Word add-in
Small firm already paying for ChatGPT PlusMove to ChatGPT Business at $25 before anything elseSame price, no training by default, admin controls
Mid-sized European firmClaude Team or EnterpriseNon-training by default; Astra for Law's index is US law; re-test Astra in Q1 2027
Large firm on Harvey or LegoraWhatever the platform routes to, plus one enterprise chat toolBoth models are already inside the platform; buy the chat tool your DMS connects to
In-house team inside a ChatGPT Enterprise companyChatGPT Enterprise, with Astra enabled by IT if you need computer useProcurement is done; add Claude for Legal plugins only if Legal has its own budget
US litigation boutiqueTrial Astra for Law against Claude plus a citation checkerThe 230M-URL index is the one feature Claude cannot match today
Anyone handling material the client would not want a third party to holdAPI with ZDR, through a vetted vendor, or no AI at allFable 5.1 cannot be run under ZDR; consumer tiers of both are out

A bake-off you can run in an hour

Leaderboards measure someone else's work. Run yours. Paste the same instruction and the same anonymised NDA into Claude (Opus 5 or Fable 5.1) and ChatGPT (GPT-6 Astra or GPT-5.6), then score both with the checklist below.

You are reviewing a mutual NDA for the Receiving Party, a mid-sized
software company incorporated in Austria. Governing law: English law.

Playbook positions:
1. Confidentiality term: 3 years from disclosure; trade secrets indefinite.
2. Residuals clause: reject. Fallback: accept only for general know-how
   retained in unaided memory, excluding source code and customer lists.
3. Non-solicitation: reject entirely.
4. Injunctive relief: accept mutual; reject any waiver of the need to
   prove damages.
5. Return or destruction: accept, with a carve-out for routine backups
   and one archival copy for legal/regulatory purposes.

Task: produce a redline as a numbered list. For each change give
(a) the clause reference, (b) the original wording, (c) the replacement
wording, (d) one sentence of rationale for our client, and (e) a
"fallback if refused" line. Flag any clause that conflicts with the
playbook but which you did not change, and say why. Do not invent clause
numbers; if the document lacks them, quote the first six words instead.
Do not cite any case law or statute.

Score each output on: every playbook point addressed; no invented text; replacement wording usable without editing; rationale a partner would sign; fallbacks sensible; conflicts flagged rather than silently changed. The contract-review use case explains how to turn the winner into a repeatable workflow, and Module 1 of the AI Lab for Lawyers course runs this comparison live so participants see the difference on their own documents.

What to do this week

  1. Check which tier your people are actually on. Clio's 2025 Legal Trends Report found 53% of firms have no AI policy or do not know of one; 34% of professionals in Thomson Reuters' 2026 survey use unsanctioned tools.
  2. If anyone is on Claude Free, Pro or Max or ChatGPT Free, Plus or Pro with client material, switch training off today and buy a business tier this month.
  3. Run the bake-off above on two real documents from your practice, anonymised.
  4. Decide on data terms, then on features. A model that is 2 points better on a benchmark and trains on your chats is the wrong model.
  5. Revisit in three months. GPT-5.5 retires from ChatGPT on 14 October 2026, Gemini 3.5 Pro is still unreleased, and Astra for Law will have its first independent tests by then.

Frequently asked questions

Is Claude better than ChatGPT for lawyers?

For drafting, redlining and long-document analysis, Claude has the edge as of September 2026: Anthropic reports Opus 5 first-turn NDA redlines nearly double Opus 4.8, and Claude models hold the top two places on the Vals Index. ChatGPT is stronger for computer-use agents and, through Astra for Law, US case-law retrieval. Both hallucinate citations and both need the same verification.

Does ChatGPT train on my legal documents?

On the Free, Plus and Pro tiers, yes, unless you switch off model improvement in Data Controls. ChatGPT Business, Enterprise, Edu and the API do not train by default, and the API keeps content for 30 days for abuse monitoring unless you have a zero-data-retention agreement. Treat every consumer tier as unsuitable for client material.

Which is safer for confidential client work, Claude or ChatGPT?

Business tiers of both are acceptable with a written policy; consumer tiers of neither are. Claude Team and Enterprise never train on content but their chat interfaces are not zero-data-retention eligible, and Fable 5.1 requires 30-day retention. ChatGPT Business and Enterprise do not train, and ZDR is available contractually. A court has held consumer Claude memos not privileged, and 20 million ChatGPT logs were ordered produced in litigation.

Which model has the lowest hallucination rate for legal work?

On the HAQQ benchmark of 3,000 answers (June 2026), GPT-5.5 had the lowest citation-hallucination rate at 3 percent but was the slowest and most expensive model; Claude Opus 4.8 scored highest overall with roughly 12 percent. Every model, including the leaders, fabricated or misapplied a citation somewhere, so the verification duty does not change with your choice of vendor.

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.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion

Tools

Claude for Lawyers: An Honest Review of Anthropic's Models for Legal Work

Claude for lawyers, reviewed as of September 2026: Fable 5.1, Opus 5 and Sonnet 5, plan prices, data retention by plan, legal plugins, benchmarks, weaknesses.

9 min read

Tools

ChatGPT for Lawyers: GPT-6 Astra, ChatGPT Work and What to Switch Off

ChatGPT for lawyers as of September 2026. Which plans train on your chats, the 20-million-log discovery order, GPT-6 Astra, ChatGPT Work and a firm setup.

8 min read

Comparisons

Claude for Legal vs Astra for Law: Anthropic and OpenAI Go Direct to Lawyers

Claude for Legal vs Astra for Law: open-source plugins and MCP connectors against a 230m-URL US law index with a ZDR API, data terms, prices, the European view.

11 min read

Guides

Legal AI Benchmarks Explained: Vals, BigLaw Bench, LAB, HAQQ and LegalBench

Legal AI benchmarks explained: what Vals, BigLaw Bench, Harvey LAB, HAQQ and LegalBench measure, the September 2026 leaders, and how to run your own evaluation.

10 min read

Guides

Which AI Tools Train on Your Data? A Confidentiality Checklist for Lawyers

Which AI tools train on your data, plan by plan as of September 2026: ChatGPT, Claude, Gemini, Copilot, legal vendors, plus bar rules and a 10-point checklist.

11 min read

Playbooks

AI Privacy Settings for Lawyers: Step-by-Step in ChatGPT, Claude, Gemini, Copilot and Perplexity

Exact steps to turn off training and check retention in ChatGPT, Claude, Gemini, Copilot and Perplexity as of September 2026, plus admin settings and a log.

10 min read