For a three to twenty person service firm, Claude and ChatGPT are closer on back-office work than either brand's marketing suggests. On the writing and summarising jobs they are the same tool. The differences live in three places: what each vendor does with what you paste, which one already sits inside the office's tools, and which one the owner keeps open. We use both at Nihul, and the prices below were checked in September 2026.
What happens to what you paste
What a vendor does with the material you put in changes at one line in the plan table. Both vendors stop training at the business tier. Anthropic's default: it does not use inputs or outputs from its commercial products to train its models, and its Team and Enterprise plans sit under that policy. OpenAI's default for business users, including ChatGPT Business, ChatGPT Enterprise, and the API, is the same: no training on inputs or outputs. If your firm sits on Claude Team or ChatGPT Business, no-training is the default, not a setting you have to hunt for.
Below that line you are relying on a setting, and the two brands start it in opposite positions. On ChatGPT's consumer plans the setting starts on: OpenAI's own words are that ChatGPT "improves by further training on the conversations people have with it, unless you opt out". On Claude's consumer plans the control is called Model Improvement, under Privacy Settings, and it starts off: Anthropic says it uses consumer chats to improve Claude only "if you choose to allow us to use your chats and coding sessions". Claude's Incognito chats are not used to improve the model even with Model Improvement switched on.
Retention is a separate question, and the answer is thinner. OpenAI offers customer-set retention windows on Enterprise, Healthcare and Edu, not on the self-serve Business plan. Neither vendor advertises zero data retention on the seats a small business would actually buy. For a small firm the practical protection is the business tier's no-training default, plus the discipline of not pasting what you would not put in an email to a client.
Exhibit 1
Both vendors stop training on your inputs by default at the business tier, and on the consumer plans the two start the same setting in opposite positions.
The prices, checked September 2026
Claude's Pro plan is $20 a month billed monthly, or $17 a month if you pay $200 up front for the year. Above Pro sits Max, which Anthropic prices "From $100 per month" for either 5x or 20x Pro's usage. Claude Team is $20 per seat per month billed annually, $25 billed monthly, and Anthropic requires a minimum of two members, with a premium Team seat at $100 per seat per month annually, $125 monthly. Claude Enterprise is quoted as "$20/seat" plus usage billed at API rates, so the seat fee is not the bill.
ChatGPT Plus is $20 a month, billed monthly. ChatGPT Pro comes in two tiers, $100 and $200 a month, differing only in usage allowance. OpenAI's team plan is now called ChatGPT Business, and it is $20 per user per month billed annually, $25 billed monthly, with a two-seat minimum. OpenAI has also added a cheaper consumer tier below Plus called Go, which its own pages say "may include ads"; the price is not printed on any page we could read.
OpenAI sells its consumer plans monthly only. There is no annual discount on Go, Plus or Pro, and annual billing is reserved for Business and Enterprise. Anthropic discounts Pro and Team for paying annually, so the annual option exists only on the Claude side. The moment the firm buys the business tier, the list price is the same on both.
Where each one sits inside the office
An assistant you have to open a second tab for is an assistant you use less. Claude in Chrome is generally available on all paid Claude plans, and Anthropic describes it as reading the page you are signed in to and then clicking, typing and filling forms. Claude also runs inside Excel, PowerPoint and Word on any paid Claude plan, with Outlook still in beta, and it connects to Gmail, Google Calendar and Google Drive on every plan, asking for approval before each action by default. On Team and Enterprise those connectors have to be switched on by an owner first.
OpenAI renamed its connectors to "apps" in December 2025. They work for all signed-in users, though some app capabilities are restricted to Plus, Pro, Business and Enterprise, and the ones the pages we read name are Google Drive, Slack, OneDrive and SharePoint. ChatGPT has a sidebar add-in for Excel and Google Sheets, installed from the Excel Add-ins menu, and it does not cover Word, which is the one gap Claude fills. OpenAI retired the Atlas browser on 9 August 2026 and folded browsing into ChatGPT itself. ChatGPT now ships a built-in browser on every consumer tier including Free.
The back-office jobs, one at a time
We have run each of these jobs on both assistants, and the everyday tasks AI can take over is where we list them. The model work is the same on these jobs; the differences are in what sits next to the work.
Drafting replies and quotes is the job we test first, because it is the one that shows the voice. We cannot find a consistent difference between the two. The drafts come out at the same level, and the real work is the editing in how to triage the inbox and draft replies and how to draft a quote from past quotes. If one of the two is already open in your tab, that one wins this job.
Summarising documents is the same story, with one real difference in the upload caps. Claude takes PDF, DOCX, CSV, XLSX and more, up to 500MB and 20 files per chat, with PDFs capped at 1,000 pages. ChatGPT accepts uploads on Free and every paid plan, but spreadsheets and CSVs are capped around 50MB, other files at 512MB, and free users get three uploads a day, while Claude does not ration uploads per day. If your summary habit is one fat PDF, the spreadsheet cap rarely bites; if it is a batch of CSVs, it does. Using AI to summarise a contract before you sign it is the version of this job we write up.
Tidying a spreadsheet export is where the add-ins matter. Claude reads CSV and spreadsheet files and runs Python over them on every plan, once code execution is switched on, and it can build an actual .xlsx and hand it back, or save it to Google Drive, on every plan including Free. ChatGPT reads .xlsx and .csv files and runs Python over them in a notebook environment, and can pull the file straight from Google Drive, OneDrive or SharePoint where connectors are enabled. QuickBooks Online and Xero both export any report to Excel, so the file is never the problem. OpenAI's own prep rules for a spreadsheet you hand an assistant are worth adopting as the house standard: descriptive headers in the first row, one row per record, plain-language column names, no stacked tables and no blank rows splitting the data.
The weekly report is the same tool again. Both will take a week of notes and a few CSVs and produce the draft, and the template you give them does the work. It is the job where sitting on the shared drive matters.
First-pass bookkeeping questions, like why this invoice is stuck or which line moved the margin, are the careful ones. The assistant can read the export and the rule and point at the candidate, but the entry stays human, and what clean books look like in a monthly rhythm is what keeps the machine work from becoming the decision.
Exhibit 2
On the five back-office jobs the two assistants are the same tool, and the differences sit in the add-ins, the browser and the upload caps.
The settings that differ under the hood
Three settings differ, and the pricing pages do not show them. Memory: Claude's memory is on by default on Free, Pro and Max, and off by default on Team and Enterprise until an owner switches it on. ChatGPT's memory is limited on Free and fuller on the paid consumer tiers, and turning a memory off with "Don't mention this again" does not delete it. Projects: Claude Projects work on the free plan, capped at five, with the larger knowledge base reserved for paid plans; ChatGPT Projects are on every plan with an unlimited project count, capping files at 5 on Free, 25 on Go and Plus, and 40 on Pro and Business.
Context windows are published differently: Anthropic prints a flat 200k for Claude, ChatGPT publishes 54K on Plus for its instant model and 256K for its reasoning model, and OpenAI warns that the share available for your own input is smaller than the headline. For back-office documents of a few dozen pages either window is a non-issue.
Which one the owner will actually keep open
We use both at Nihul, and the one that gets used on a given job is the one already open where the work sits. Inside Word that is Claude, because ChatGPT's add-in does not cover Word. In Excel, Google Sheets or a browser tab it is whichever one you opened first, and habit decides more of this than any feature table.
The verdict by firm shape
A solo owner cannot buy the business tier as a single seat, because Claude Team requires a minimum of two members and ChatGPT Business has a two-seat minimum. You stay on a consumer plan and manage the setting yourself: on ChatGPT that means opting out of training, on Claude the Model Improvement control already starts in the off position. If pasting client material is the norm, the cleanest move is the two-seat business plan of the assistant the office already lives in. If you stay solo, Claude's Pro at $17 a month on annual billing beats ChatGPT Plus at $20 with no annual option, with the cheaper Go tier as the unpriced wildcard.
A five-person office buys the business tier. The no-training default is the guarantee, and the list price is the same on both at $20 per seat per month billed annually. The choice is made on the office suite: if the office lives in Word, Claude covers it and ChatGPT's add-in does not; if it lives in Excel, Google Sheets and a browser, both cover the sheet and ChatGPT's built-in browser is already open. On Team, Claude's memory is off by default until an owner switches it on, and the Gmail, Calendar and Drive connectors need the same owner switch, with approval before each action by default.
A twenty-person firm with a shared drive chooses on the connectors, because the drive is where the back office lives. Claude's Gmail, Google Calendar and Google Drive connectors run on every plan and ask for approval before each action, but on Team they sit off until an owner turns them on. ChatGPT's apps cover Google Drive, OneDrive and SharePoint for the signed-in users, with some capabilities restricted to the paid tiers. Whichever assistant the drive and the inbox already sit beside is the one to standardise on, and the other one stays as the spare for the jobs the first one's add-ins cannot reach.
Exhibit 3
A solo owner stays on a consumer plan, a five-person office buys the business tier, and a twenty-person firm follows its shared drive.
If the question has grown past choosing a plan, if it is which of the firm's jobs should touch an assistant at all and on which tier, that is part of the work we do under AI enablement.