Claude vs ChatGPT vs Gemini for Architects: Which to Pay For

Pay for one general assistant seat, about $20 a month, for each person who writes briefs, letters or specifications, and pick between Claude, ChatGPT and Gemini by testing your own three jobs. If the practice already uses Google Workspace Business Standard or above, Gemini is included, so test it first and buy a paid seat only where it falls short.

The price gap is small, so the choice turns on three things you can check in an afternoon: how each tool handles your documents, what it does with client data, and where your writing already lives. The test below uses invented project material, so nothing confidential needs to leave the building.

Follow me on Instagram@sagnikteaches

What a seat costs in each product

These are USD list prices from each vendor's pricing page in October 2026. Team and Business plans need at least two seats. Google's AI Pro plan is shown for completeness, but Google's FAQ says it needs a personal Google Account that you manage yourself, which is a poor fit for client files.

Connect on LinkedInSagnik Bhattacharya

Workspace Starter, at about $7, includes Gemini in Gmail and the Gemini app but not in Docs, Sheets or Drive, so the Standard price is the one to compare. If the Workspace pricing page shows an introductory discount, compare on the standard price, which is what you pay once the offer ends. Claude's Max plans (from $100) and ChatGPT Pro (from $100) are built for heavy individual use, which a practice's writing seldom needs.

Subscribe on YouTube@sagnikteaches

Put the cost against time. At $75 an hour of internal cost, a $20 seat is covered by 16 minutes of saved time a month ($20 ÷ $75 × 60). The seat price will never decide this; the quality of the first draft and the safety of the data will. For a wider look at what AI costs a practice, see how much AI costs a small architecture practice each month.

Annual billing saves $40 a year on a Claude Pro seat ($200 against $240) and $60 a year on each Team or Business seat ($240 against $300). For two seats that is $120 a year, which is real money, but it commits you for twelve months, and seat cuts happen only at renewal. Test on monthly billing first and switch to annual once the choice has survived a quarter.

File limits that matter for drawings and specifications

Architects feed these tools long PDFs, so file handling matters as much as the writing. The figures below come from each vendor's help pages:

The Claude row is the one that bites. A 140-sheet drawing set uploaded as one PDF would be read as text only, so notes and schedules would be understood but the graphics would not. Split a set into runs of 100 pages or fewer if you want it to read plans and sections. None of the three tools replaces your drawing software, and none should be trusted to read a dimension off a scaled drawing; use them for the words around the drawings.

Where the writing lives matters as much as the model

A tool that writes inside the application you already use saves the copy-and-paste step, and on short jobs that step is a large share of the time. Google says Gemini works in Gmail, Docs, Sheets, Slides, Drive and Meet on Workspace Business Standard and above, so a client letter drafted in Docs needs no paste at all. ChatGPT Business links to other services through apps (formerly connectors), which are on by default and which an admin can switch off, and Claude's pricing page lists connectors on Pro and Team, with Claude in Chrome and Microsoft 365 on Pro. Integrations change often, so write down the two applications where your practice does most of its writing and check what each vendor currently offers there before counting it as a reason to buy.

Specifications are the exception. They usually live in specification software or a master document that none of these assistants writes into directly, so expect to paste clauses in whichever you choose. For that job, file handling and accuracy decide, not integration. A useful split is to choose the writing assistant on the brief and letter tests and to treat the specification test as a veto: any assistant that invents standards fails, whatever its other scores.

Check the data settings before you upload anything

Run this check on each product before the test, and again before anyone uses it on real work:

• Training. OpenAI's Business page says "no training on your business data by default", and Claude's pricing page says no model training on your content by default for Team. On personal plans (Free, Pro, Plus, Max), training is a setting each user can switch off, so someone has to do it on every account. Gemini in Workspace business plans also does not train on business content by default; personal accounts have their own activity setting.
• Who can see chats. On team plans, assume the account owner or admin can view and export members' conversations, and tell your staff so.
• Retention. Deleted chats are not necessarily gone at once. OpenAI says deleted ChatGPT chats are removed within 30 days, and Temporary Chat can be kept for up to 30 days for safety. Check each vendor's current page for the retention period that applies to your plan.
• Client terms. Read the confidentiality clauses in your appointment letters and any NDA. Some clients restrict third-party processing of project information. The tutorial on copyright and client-data risks for architecture practices covers the questions to put to clients.

For the test itself, use an invented project. That removes the data question while you decide, and the settings above only need to be right before you move to real files. The AI policy checklist for architects shows how to record the decision.

The test: three jobs, one afternoon

Run the same inputs through each assistant in a fresh chat, twice each, so that one lucky or unlucky answer does not decide. Time yourself from pasting the prompt to a version you would send. Save every output in a folder, strip the vendor names, and ask a colleague to score them blind. One rule applies to all three jobs: any standard number, policy reference or manufacturer you did not paste in is an automatic fail for that run, because a confident invention is worse than a gap.

To keep costs down, screen on the free tiers first, then pay monthly (not annually) for the two that scored best. Three monthly plans cost about $60 in total, and two cost about $40, so the whole test is cheaper than one hour of a principal's time.

Job 1: the client brief

Paste this invented email and ask for a one-page design brief, ten questions for the client and a room data list.

We are a family of four with a detached house from the 1930s and about
$180,000 to spend. We want the kitchen and dining room opened into one
space with doors to the garden, a utility room, a ground-floor study that
can become a bedroom for my mother, and a bigger upstairs bathroom. We want
to start in spring and stay in the house during the work if possible. We
like natural materials, and my husband works from home two days a week.
The garden has a mature oak we want to keep.

Task: (1) One-page design brief with headings Objectives, Rooms and areas,
Budget, Programme, Constraints, Unknowns. (2) Ten questions for the client.
(3) A room data list. Do not invent dimensions, costs or consents.

Score it on whether the brief keeps every stated constraint (the oak, the accessible ground-floor bedroom, living in the house during works, home working), whether the Unknowns heading lists real gaps such as consents, structure and services, and whether any dimensions or costs appear that you did not provide. A weak output says things like "kitchen-diner of 28 square metres" and "estimated cost per square metre"; both are inventions. Fixing one takes ten minutes, and it is the reason to read every figure in a draft.

The Unknowns heading is where the assistants differ most. A weak draft might read: "Unknowns: final budget; planning approval; start date; choice of kitchen supplier." Three of those four restate the email rather than find its gaps. What the email cannot tell you is whether the oak's roots limit how deep the extension can go, what the existing structure and services allow once walls come out, whether the ground-floor bedroom needs step-free access to a bathroom, and whether a family of four can really live in the house during the works. A strong draft raises questions like those without being asked. When a draft does not, add them, and mark down the assistant that missed them, because finding gaps is the skill you are paying for.

Job 2: the planning letter

This is the cover letter that goes with a planning or permit application. The test is whether the assistant sticks to the facts you supply. Compare two prompts built on the same notes:

Weak: Write a planning cover letter for a rear extension to a detached
house. Make it persuasive and cite relevant policies.

Better: Write a one-page cover letter to the approving authority.
Facts: single-storey rear extension, 4.2 m deep, flat green roof; rear
windows replaced; mature oak retained with root protection measured by
the arboriculturalist; neighbours told on [date]. Policy extracts you may
cite are pasted below and nothing else: [paste]. If a fact is missing,
write [TO CONFIRM]. Do not invent policy numbers, dates or consultations.

The weak prompt invites policy numbers that sound right and are not. The better one produces a letter with square-bracket gaps you can fill. Score on whether every claim traces to your facts, whether the tone is calm, and how many minutes it takes to make the letter ready to sign. The tutorial on where to draw the line on planning applications goes deeper on what AI should not write.

Job 3: the specification clause

Ask each assistant to draft a clause for fire stopping to service penetrations in a two-storey extension, with headings General, Products, Execution and Submittals. Add this instruction: "Where a standard, rating or manufacturer is needed and I have not supplied one, write a placeholder in square brackets."

A good draft reads like a clause you could issue after filling the brackets: scope, performance requirement, installation, inspection, certificates, with every number or standard left as a placeholder. A bad draft cites a numbered standard and a 120-minute rating you never gave it, in the confident tone of a finished clause. That draft is dangerous precisely because it looks complete. Also check internal consistency: a clause that asks for a product certificate under Submittals but never mentions it under Execution has a gap a contractor will exploit. The tutorial on using Claude for specifications and planning reports includes more spec prompts.

Job 4, optional: a long consultant report

If your practice reads long structural, energy or survey reports, add a fourth run. Take an invented or sanitised 60-page report, upload it and ask: "List every action in this report with the owner, the deadline if stated, and the page number." This job separates the products on file handling rather than writing. Spot-check five of the page references against the PDF. A tool that returns plausible actions with wrong page numbers is guessing, and you now know to treat its references as hints, not citations.

Then repeat the run with the same report split into two files. If the answers change, your result depends on how you upload, which is worth knowing before you rely on it. Remember the limits: Claude reads a PDF over 100 pages as text only, so diagrams and tables set as images are missed; ChatGPT caps text and document files at 2 million tokens each; and the Gemini app takes up to 10 files per prompt.

Scoring the three runs and turning minutes into money

Score each run 0 to 2 on four things: fidelity (nothing invented), completeness (every constraint kept), tone and format, and edit time. Add the minutes you spent from paste to send-ready. Here is how the arithmetic works, using made-up scores for three anonymous assistants; your own runs decide.

Assistant A wins on score and B on speed, with the difference between them small, so the tie-break is data settings and where the writing lives. Assistant C saves only 35 minutes across the three jobs and had the most invented content, so it drops out. Now turn the minutes into money. Say the practice produces six briefs, eight letters and ten clauses a month. Assistant A saves 30, 20 and 10 minutes a job: 6 × 30 + 8 × 20 + 10 × 10 = 440 minutes, about 7.3 hours. At $75 an hour that is roughly $550 a month against seats of $40 to $50 for two people. Notice that the large number comes from the writing volume, not from the vendor's name.

Run each test twice and average, because the same prompt can produce different results on different days. And rerun it each quarter: these products change monthly, and the one that lost in one quarter can win the next. The Gemini or ChatGPT for a Google Workspace business tutorial helps if your result is close and your team lives in Gmail and Docs.

Seats by person: who gets a paid assistant

Match the seat to the person, not the practice. Take a five-person practice on Google Workspace Business Standard, with two architects who write most documents, two technologists and one administrator:

That stack adds $40 to $50 a month to the Workspace bill. If the practice runs on Microsoft 365 instead, the equivalent first test is the Copilot Chat that comes with its business plans, with the same two paid seats for the heavy writers. If Copilot Chat falls short inside Word and Outlook, the paid step is Microsoft 365 Copilot Business at $21 a user a month on annual billing ($25.20 billed monthly), with an $18 promotional price for existing Microsoft 365 business customers until 31 December 2026. Run it through the same three jobs before you buy it for anyone. The wider four-way comparison linked at the end covers that route.

Setting up the winner so the whole practice benefits

A tool that one person prompts well and everyone else prompts badly wastes most of its value. Put the practice's standing instructions in a shared Project wherever the plan allows it. OpenAI says project sharing works on every ChatGPT plan, with Business adding workspace sharing, and Claude's Team plan includes projects with sharing and collaboration. Start with something like this and refine it each time a draft goes wrong:

You assist a small architecture practice. Voice: calm, precise, plain English.
Never cite a standard, policy number, rating or manufacturer unless the user
has pasted it in this conversation. Where one is needed, write [STANDARD],
[POLICY] or [MANUFACTURER] and list the gaps at the end.
Never invent dimensions, areas, costs, dates or consultation history.
If a fact is missing, write [TO CONFIRM].
Do not describe our design as compliant with anything. Say "to be checked
by the architect".
Show arithmetic line by line.

Add the practice's standard paragraphs (terms, scope exclusions), a template for each of the three test jobs, and a one-page "never paste" list covering client contact details, bank details and anything under NDA until your policy says otherwise. The bracketed placeholders do the quality control for you: a draft with no brackets in a document that needs a standard is a draft to distrust.

Ninety days on: three checks before the seats renew

A choice made in an afternoon needs checking against real work. After a quarter, re-time five recent documents of each type against the minutes in your test, count the invented items you caught per ten drafts (keep a tally in the shared project's notes), and ask each seat holder which jobs they used it for. A filled-in check for the five-person practice might read: briefs down from 60 to about 35 minutes, letters from 40 to 25, two invented policy references caught in 30 drafts, the principal using the seat daily and the project architect twice a month. The minutes justify the principal's seat. The project architect's seat is the one to question: either specification work moves into it with the project's standing instructions, or it goes at renewal. And two inventions in 30 drafts is a reminder that the placeholder rule is doing real work, so keep it in the instructions.

Three mistakes that waste a practice's money

Five seats for five people. Two seats that get used every day beat five that get opened twice. Check usage after 30 days and cut at renewal.

Real client files in a free test. It is easy to start the test with a live project because it is to hand. Use the invented brief instead until the settings are checked and the plan is chosen.

Choosing on one impressive output. One brilliant run says little. Scores across three jobs, run twice, blind-scored, do. If two assistants come out within a couple of points, choose the one whose data settings and file handling fit the practice, and revisit in a quarter.

Questions practice owners ask about paying for AI

Can these assistants open Revit or DWG files?

Claude's help page lists PDF, Word, spreadsheet, text and common image formats, not CAD files, and I found no CAD list on OpenAI's or Google's pages. Export PDFs or images from your drawing software. If you want Claude to read the graphics, keep each PDF to 100 pages or fewer, because longer files are read as text only.

Should every person in the practice get a seat?

No. Buy seats for the people who write briefs, letters, specifications and reports, and test whether anyone else would use one weekly. A five-person practice where two people do most of the writing needs two seats, not five. Review usage after 30 days and drop the seats nobody opened at renewal.

Is a personal Pro or Plus plan acceptable for client work?

It works, but it leaves you without admin control, shared projects or a default no-training setting. Consumer plans let each user switch training off, and each person must remember to do it. Team and Business plans start at two seats and do not train on your content by default, which suits client files better.

Do I need to tell clients I use AI?

Check your appointment terms and any confidentiality clause first, because some clients restrict which third parties can handle their project information. Rules on disclosure vary by contract and by client. Ask your professional indemnity insurer or solicitor how they want AI use recorded, and write the answer into the practice's AI policy.

Further reads

• Is Claude Team Worth It for a Small Business? — Whether Claude Team's two-seat minimum pays off.
• Is ChatGPT Plus Worth It for an Architect? A Practice Cost Check — A practice-level cost check on ChatGPT Plus.
• ChatGPT vs Claude vs Gemini vs Copilot for Small Business (2026) — The same three assistants plus Copilot, for any small firm.
• How Architects Use AI for Fee Proposals, Briefs and Paperwork — AI for fee proposals, briefs and paperwork.
• AI Training for Architects: What a Practice Workshop Covers and Costs — A half-day AI workshop agenda for an architecture practice, modules by role, and the cost per person once staff time is counted, set against self-taught routes.
• AI Marketing for Architects: Case Studies, Website Copy and LinkedIn — Turn a site-visit voice note into a case study, service-page copy and LinkedIn posts that sound like your practice, with permission and claims checks.
• AI Programs for Architects: Room Data Sheets From a Client Brief — A worked example of turning a messy client brief into room data sheets with AI, and the review pass that catches what the draft invents or drops.
• AI Software for Architectural Drawings: Notes, Schedules and Sheet Checks — How a small practice can check a drawing set for index, note and schedule errors with ChatGPT or Claude, and what the architect still signs.
• AI Tools for Architects 2026: A Toolkit From Brief to Handover — Which AI tool to use at concept, design, tender, site and handover, what each costs on its own pricing page, and three stacks for a five-person practice.
• AI for Architects: Draft Design Schematics From a Client Brief — A brief-to-schematic workflow for small practices: stress-test the brief, draft an area schedule and options, then run the checks that stay with the architect.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Claude pricing page and help article on uploads and PDF handling (updated July 2026); ChatGPT pricing page and OpenAI's ChatGPT Business pricing page; OpenAI's file uploads help article; Google's Gemini Apps upload help and Gemini Notebook sources help; Google AI plans and Workspace pricing pages; Microsoft 365 Copilot Business pricing, all October 2026. Test scores, timings and job volumes are illustrative.