Claude or ChatGPT for a Law Firm: Which One to Pay For

Pay for a team plan on whichever one passes a three-task trial on anonymised copies of your own documents, because the prices match: $20 a seat a month billed annually, $25 monthly, two seats minimum on both. Neither team plan trains on your content by default. Neither removes your duty to check every clause and authority yourself.

What differs is everything around the price: how long a document each tool will take, what an owner can see and control, and whether you can get audit logs without a large-firm contract. Claude's Enterprise plan has a 20-seat minimum, so a five-lawyer firm that wants audit logs there pays for 20 seats, at least $400 a month before usage. ChatGPT's Enterprise plan is by quotation. Everything else a small firm needs sits on the cheaper team tiers.

Follow me on Instagram@sagnikteaches

What a four-seat firm pays on each platform

List prices below come from each vendor's pricing page as of 1 November 2026. Both change plans often, so confirm before you commit.

Connect on LinkedInSagnik Bhattacharya

The prices are identical, which is why price should not decide this. Take a firm with three lawyers and one paralegal. Four Standard seats cost $80 a month billed annually ($960 a year) or $100 a month billed monthly ($1,200 a year). Upgrade the busiest lawyer to a Premium seat and the monthly bill on annual terms becomes 3 x $20 + $100 = $160. The tool earns that back quickly: at an assumed billing rate of $200 an hour, a $25 seat is repaid by 7.5 minutes of saved work a month. The cost that matters is the time you spend checking output, not the subscription.

Subscribe on YouTube@sagnikteaches

Start on monthly billing for the first quarter. Both vendors' help pages say a seat reduction takes effect only at the next renewal or billing cycle, with no refund for unused seats, so an annual commitment for four seats that turns out to need three is money you cannot reclaim. Adding seats is easier: both allow it mid-term, and OpenAI charges added seats pro rata straight away.

Enterprise is where the two diverge on price. Claude Enterprise is listed at $20 a seat a month billed annually with a 20-seat minimum, plus usage at API rates, so the floor is $400 a month before usage. ChatGPT Enterprise is custom-priced through sales. Both add audit logs, custom retention and finer access controls. If nobody is asking you for those, you are not shopping at that tier. For a seat-cost breakdown aimed at another professional firm, the consultancy seat-cost comparison uses the same list prices.

Three legal tasks, same inputs, scored the same way

Pick tasks that mirror the firm's real week. These three cover drafting, review and long-document work, and each can be run on public or anonymised material so no client is exposed during the comparison. For a fuller same-task protocol, the lawyers' same-task test walks through running it on your own matters. The version here is built for a buying decision rather than a general evaluation.

1. A client letter. Give both tools the same bullet instructions (the facts, the advice you want conveyed, the deadline) and ask for a first draft. Score it on whether every fact traces to your bullets, whether it states any law you did not give it, and how many minutes you spend fixing it.
2. A clause review. Use a 40-page commercial agreement with six problems you planted on purpose, and a six-point playbook. Score issues found, and separately score invented or wrong clause references, because a confident wrong reference costs more than a missed issue.
3. A bundle chronology. Use about 200 pages of emails, letters and statements and an answer key of 40 dated events. Ask for a chronology with the PDF page of each entry. Score correct entries and wrong page references.

The clause-review prompt is where firms usually under-specify. A version with the playbook written in:

You are assisting a lawyer. Review the attached supply agreement against this playbook.
For each of the six points: quote the clause number and the exact wording, say whether it Meets, Partly meets or Fails the playbook, and suggest fallback wording.
Do not rely on general knowledge of the law. If the document does not address a point, write "Not addressed". Never invent a clause number.

Playbook:
1. Renewal: no automatic renewal term longer than 12 months; notice period no more than 60 days.
2. Liability: cap is mutual and at least 12 months of fees.
3. Termination for convenience: either side, on 60 days' written notice.
4. Payment: 30 days from invoice; no unilateral price increases.
5. Confidentiality: mutual, surviving 3 years after termination.
6. Disputes: mediation before any court action.

An illustrative row from the kind of table you should get back, with the check you must do:

Point 1, Renewal. Clause 14.2: "This Agreement renews automatically for successive terms of 24 months unless either party gives 120 days' written notice." Fails: renewal term is 24 months and notice is 120 days. Fallback: "renews for successive terms of 12 months unless either party gives 60 days' notice."

That row looks perfect, which is the danger. Open the agreement and confirm that clause 14.2 exists, that it says 24 months, and that the quotation is verbatim. A tool that paraphrases while claiming to quote, or cites 14.2 when the clause is 14.3, has failed the part of the task a supervising lawyer cannot skim.

Long documents: what each tool takes, and the page-187 test

There are two different limits, and vendors' pages mix them. One is how big a file you can attach. The other is how much the model can hold in view when it answers. A bundle can be accepted by the first limit and still overflow the second.

The ChatGPT figures in the last row come from the comparison table on OpenAI's business pricing page, which describes them as approximate because the system uses part of the window for instructions and reasoning. Claude's plan table says its window goes up to a million tokens, depending on the model. Neither number guarantees that every page of a 600-page bundle is weighed equally in a single answer, so test it.

The test: plant one distinctive line deep in the bundle, say an email on page 187 of 200 recording that payment terms were extended to 60 days. Then ask a question whose correct answer depends on that email: "What payment terms applied after the March correspondence, and where is that stated?" A tool that answers from the contract's original 30 days has not used page 187. A tool that gets the terms right but cites the wrong page has read it loosely. Run it three times with the needle in different places, because results vary between runs.

Two preparation points save arguments later. First, scanned bundles: run text recognition in your own PDF software before uploading and confirm you can select the text, because neither vendor's page promises that image-only scans are read reliably. Second, redactions: a black box drawn over text in some PDF editors leaves the words underneath. If you send redacted material to any assistant, flatten the file first, or you may have sent exactly what you redacted.

Privilege, confidentiality and what the defaults leave open

Business plans from both vendors say they do not train on your content by default, and both vendors' consumer plans have a setting you can switch. That answers the training question. It does not settle confidentiality, and several smaller details matter more to a firm.

• Retention is separate from training. Anthropic says chats on its commercial products stay in the product until you delete them, and that chats flagged as breaching its usage policy can be kept for up to two years, with trust and safety classification scores for up to seven. Ask OpenAI the equivalent question and get the answer in writing.
• Departing staff. OpenAI's help centre says that on ChatGPT Business a removed member loses access at once, owners and admins cannot see members' private chats by default, and members cannot export their own workspace data; its privacy FAQ describes broader admin access, so get the answer in writing. Treat every chat as scratch paper, and save the final work product to your document management system the same day.
• Browser agents. Anthropic's safety page for Claude in Chrome says to avoid legal documents and contracts, among other sensitive material. Keep browser-driving features away from live matters on any vendor until your risk policy says otherwise.
• Hostile documents. Anthropic's help article on code execution warns that instructions hidden in external files or websites can trick the model into leaking what is in its context. An opposing party's disclosure is an untrusted file. Do not load it into the same Project as your own client's privileged notes.
• Your own rules. Professional conduct rules, insurer wording and outside-counsel guidelines from corporate clients decide what you may upload and whether you must tell the client. Read them before the vendor's settings pages, because the rules bind you and the settings do not.

Claude for Legal plugins: what May 2026 added, and what you need

In May 2026 Anthropic announced Claude for Legal: more than 20 connectors to legal software and 12 practice-area plugins, among them commercial, corporate, employment, privacy, regulatory, intellectual property and litigation. Its own announcement says each plugin opens with a short setup interview that learns your playbook, escalation chain, risk calibration and house style, so answers are tuned to your team rather than generic. It also describes Claude working inside Word, where skills handle drafting and redlining against your playbook with tracked changes, and inside Outlook, Excel and PowerPoint.

Read the availability lines before you plan around it. The announcement says the plugins and connectors are open source and available in Claude Cowork, which Anthropic began merging into Claude itself in September 2026 and has been rolling out plan by plan since, and that Enterprise admins can enable them in workspace settings. A later help article says the single Plugins page began rolling out to Free, Pro, Max, Ultra and Team plans on 29 October 2026, with general availability on 9 November, and Enterprise on 30 November. On Team, an owner has to switch on Skills and the cloud code execution and file creation setting first, because plugins need both. Check your plan's current state on the day you buy, not a summary written months earlier, this one included.

The connectors named in the announcement include document and matter systems such as iManage, NetDocuments, Box and Docusign, research and review platforms such as Thomson Reuters CoCounsel, Everlaw and Relativity, and the Harvey assistant. Whether a connector helps you depends on whether your firm uses that system. A three-lawyer firm on a plain shared drive gets little from a connector to a document system it does not own. The plugin part, by contrast, works for any firm that can write down its playbook, and the setup interview is a decent prompt for doing that.

On the ChatGPT side, the business pricing page lists plugin creation and sharing, a skills beta, and extensions for Excel, Word, PowerPoint and Google Sheets on Business. I did not find a practice-area bundle published by OpenAI to match Anthropic's legal set, so check its plugin directory when you decide. A fair summary: Claude has a vendor-published legal suite worth a trial; ChatGPT has the general building blocks and the same Projects feature, and a firm could assemble something similar from them with more effort.

Treat the quotations on Anthropic's legal pages, including benchmark claims and law-firm testimonials, as marketing. Your own scored trial is the evidence. A plugin also does not make an output correct. It makes the output follow your playbook, which is useful only if the playbook is good. If you would rather compare with purpose-built contract tools, the contract review software comparison sets two of them side by side.

A scored purchase for a four-seat firm, with two different answers

To show how the weighting works, here is a hypothetical trial. The scores are invented for illustration and say nothing about how either product performs; "Tool A" and "Tool B" could be either. Suppose a firm runs it on public contracts and anonymised letters using individual plans (about $40 for the pair of subscriptions) and records the following.

Convert the letter task by dividing the best time by each tool's time: Tool B scores 100 per cent (14 over 14) and Tool A 77.8 per cent (14 over 18). Now weight by what the firm does. A commercial practice that spends half its time on contract review, 30 per cent on letters and 20 per cent on bundles gets Tool A 0.5 x 66.7 + 0.3 x 77.8 + 0.2 x 82.5 = 73.2, and Tool B 0.5 x 83.3 + 0.3 x 100 + 0.2 x 62.5 = 84.2. Tool B comes out ahead, even though Tool A scored higher on bundles. A litigation firm with 20 per cent on contracts, 20 per cent on letters and 60 per cent on bundles scores Tool A at 78.4 and Tool B at 74.2, and the answer flips. Same data, different firm, different purchase.

Two refinements keep the exercise honest. Repeat the clause review with a different agreement before trusting a gap of a few points, because one document can flatter either tool. And decide in advance what margin counts as a result: if the weighted totals are within about five points, choose on administration instead, meaning single sign-on, whether your document system has a connector, and how easily the owner can see who has the tool.

House rules to write before anyone pastes a client name

A one-page policy beats good intentions. The short version below can be pasted into your own wording and agreed by the partners.

1. Only the firm's paid team accounts are used for client material. Personal accounts are for nothing that identifies a client or matter.
2. Where clients' contracts or outside-counsel guidelines restrict AI, the restriction wins, and the fee earner notes it on the matter file.
3. Every authority, quotation and clause reference is checked against the primary source by the person who signs the work.
4. The final version goes into the document management system the same day. Chats are not records.
5. Documents from the other side are loaded only into a scratch chat, never into a shared Project.
6. If client material goes into the wrong account, the person deletes the chat, tells the supervising partner within a day, and the partner decides whether the client must be told.

The firm's engagement letter may need a sentence on AI tools too. Wording depends on your professional rules, so write it with whoever handles your risk and compliance. If you want a structured way to compare candidate tools beyond these two, the two-week trial scorecard and the vendor questions for tools that touch client data cover what to ask.

A way to settle it that you can explain to a partner or an insurer

Run the three tasks on public or anonymised material, on individual plans, over a week. Score as above and weight by your own work mix. If the weighted totals are within about five points, choose on administration and cost. Buy at least two seats on the higher scorer, billed monthly, for a quarter; add a Premium seat only for someone who hits usage limits more than once a week. Move to Enterprise only when a client, regulator or insurer asks for audit logs or retention control, since that is where the real price difference sits. Repeat the trial every six months, and sooner if either vendor changes plan terms, because both do so often.

Keep one expectation fixed throughout. These tools produce drafts and readings, not advice. The lawyer who signs owns every sentence, whichever tool wrote the first version.

What partners ask before buying seats

Can a small firm use the free plans for client documents?

Not sensibly. Free and individual plans have a setting that decides whether chats can be used to improve the models, and a busy team will forget to check it. Team and business plans do not train on your content by default and give an owner admin controls, which is the minimum for client material.

Does a small firm need an Enterprise plan?

Only if an insurer, a client or your own risk policy demands audit logs, custom retention or role-based access. Both vendors list those features under Enterprise. Claude Enterprise has a 20-seat minimum, so a five-lawyer firm usually stays on a team plan and keeps matter records in its own document system.

Will either tool cite real cases?

Neither can be trusted to. General-purpose assistants can produce plausible authorities that do not exist, so every case, statute and quotation must be checked against a primary source before it goes into advice or a filing. Connecting a legal research service helps, but you still verify.

Should we trial both tools on live client matters?

No. Build the trial from public contracts or documents with the names and figures changed, so nothing confidential is exposed while you compare. Move to client material only after you have chosen a plan, set the privacy options and written the house rules.

Further reads

• Per-Seat or Pay-As-You-Go? Legal AI Pricing for Small Firms — Per-seat versus pay-as-you-go pricing for legal AI tools.
• ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — When a purpose-built legal product beats a general assistant.
• AI Paralegal Software: What a Small Firm Should Test Before Paying — What to test before paying for paralegal software.
• AI Implementation Plan for a Small Law Firm: The First 90 Days — A first-90-days plan once you have chosen a tool.
• Can Lawyers Use AI Note Takers in Client Meetings? — Recording and note-taking rules for client meetings.
• AI Automation Agency for Law Firms: Questions Before You Sign — Questions to ask before hiring an outside agency.
• ChatGPT Business for Law Firms: Seats, Privacy Settings and Cost — ChatGPT Business cost and privacy settings for a small firm in depth.
• AI for Immigration Lawyers: Docketwise or ChatGPT for Forms? — Run one immigration form through Docketwise and ChatGPT: what each catches, the cost per filing, the privacy settings and the steps that stay manual.
• Free ChatGPT for Lawyers: Where the Free Plan Stops Being Safe — Free ChatGPT suits generic legal drafting, not client facts. What the Free plan keeps, how to anonymise a matter, and when a business plan becomes necessary.
• AI Discovery Review for Defence Firms: Index, Search and Cross-Check — Index discovery, map each statement to the disputed issues, build a contradiction table with page cites, then verify every cite by hand before filing.
• AI Legal Brief Writing: A Draft and Cite-Check Routine — A routine for small litigation firms: build an authority table, draft only from it, cite-check each source line by line, and meet court AI-disclosure rules.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: claude.com/pricing, openai.com/business/pricing, Anthropic's May 2026 announcement for the legal industry, Claude plugins help article, Claude and OpenAI file-upload help pages, Anthropic and OpenAI privacy pages (checked 1 November 2026).