Claude vs. ChatGPT vs. Gemini: choosing an assistant

Editor's note: This is a documentation-led edition. We have not yet run our own hands-on tests of Claude, ChatGPT or Gemini. What follows rests on official documentation, published pricing and independent studies we read in full, all accessed on 16 October 2026. A tested update will follow.

Most people pick an AI assistant on reputation. But the differences you can check before paying are mostly in the plan sheet: what each tier costs, how much text it can read at once, how big a file it accepts, whether your chats can be used for training, and whether it works inside the software you already use.

All three products run on large language models and change often. Model names, limits and prices in this piece are dated and come from each vendor's own pages. This is the hub for our assistant coverage. The single-product pieces, our ChatGPT review and Claude review, go deeper on each.

What each one is for

ChatGPT (OpenAI) has the broadest feature list: image creation, voice with video, deep research, scheduled tasks, an agent workspace called ChatGPT Work and the Codex coding agent.[1] Free and Go users chat with GPT-6 Luna. Plus adds GPT-6 Sol, GPT-6.1 Sol and GPT-6 Astra.[1]

Claude (Anthropic) leans towards documents, writing and code. Every plan includes projects, file creation, web search and connectors; paid plans add Claude Code and a Research mode.[2] Free users get the Sonnet and Haiku models; Opus starts at Pro.[2]

Gemini (Google) is tied to an ecosystem: paid plans put it inside Gmail and other Google apps and bundle cloud storage. Even the free tier includes Deep Research.[3] The free plan uses Google's Flash-Lite model; AI Pro adds Gemini 4 Argon and Google's Pro model.[3]

Side by side: the documented facts

US list prices before tax, checked 16 October 2026; sources in the sections below. A million tokens is roughly 1,500 pages of text, by Google's estimate.[4]

Plans and prices

The table gives the headline prices. Four details are easy to miss.

ChatGPT Go's price is not on the live page. The pricing page did not show a dollar figure for Go when we checked. The $8 a month figure is the US price from OpenAI's January 2026 launch post, which says Go is priced locally in some markets.[5] The pricing page adds that Go "may include ads".[1] Plus is $20 a month, with no annual option.[6]

Pro tiers differ in more than usage. ChatGPT Pro comes at $100, $200 or $500 a month, and only the $500 tier includes the faster "Ultrafast" mode.[7] Claude Max starts at $100 for five times Pro's usage, with a 20x option.[2] Google AI Ultra is $99.99 or $199.99, for five or twenty times AI Pro's limits.[3]

Claude Enterprise has no fixed all-in price. It costs $20 per seat a month, billed yearly with a 20-seat minimum, plus usage charged at API rates.[2] ChatGPT Business and Claude Team both offer Premium seats at $125 a month, or $100 yearly.[8][2]

Gemini can come with software you already pay for. Google Workspace Business Starter ($7 a user a month) includes basic Gemini app access; Standard ($14) and Plus ($22) get expanded access.[9] Paid personal plans also bundle Google One storage, from 400 GB to at least 20 TB.[3]

Context windows and file limits

The context window is how much text a model can work with in one go. Here the three diverge most, and app figures differ from the API figures vendors advertise.

• ChatGPT publishes two sets of numbers. The fast "Instant" mode has 27K tokens on Free, 54K on Go and Plus and 128K on Pro. Reasoning models get 256K on Go and Plus and 400K on Pro.[1] OpenAI adds that the space for your input is "smaller than the total window", because instructions, memories and reasoning share it.[1]
• Claude's pricing page lists a context window of "up to 1M, varies by model" on every plan.[2] Its help centre lists 1M tokens for the newest models on paid plans (Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1) and 200K for models outside its table.[10]
• Gemini sets context by plan: 32K tokens without a paid plan, 128K on AI Plus and 1 million on AI Pro and Ultra.[4]

File limits differ too. ChatGPT accepts documents up to 512 MB, capped at 2 million tokens per text file, and up to 80 files every three hours; Free users get three uploads a day.[11] Claude accepts up to 20 files per chat at 500 MB each, and PDFs up to 1,000 pages. It reads charts and images only in PDFs of 100 pages or fewer.[12] Gemini takes up to 10 files per prompt, 100 MB each or 2 GB for video.[13]

A big window shows what fits, not what the model reliably uses. Our Claude vs. ChatGPT comparison for long documents takes this further.

Data, training and retention

ChatGPT. On personal plans, conversations can be used for training unless you turn off "Improve the model for everyone". Temporary chats are not used for training and may be kept for up to 30 days for safety.[14] One catch: if you give a thumbs up or down after opting out, OpenAI says the whole conversation "may be used to train OpenAI models".[14] OpenAI does not train on Business, Enterprise or Edu workspace content by default.[14][8]

Claude. Anthropic's privacy centre says consumer chats are used for training if you allow it or if a chat is flagged for safety review; incognito chats never are.[15] Its pricing page labels consumer training "Opt-out", so check the setting rather than assume.[2] If you allow training, Anthropic may keep de-identified chats for up to five years; deleted chats go from its systems within 30 days.[16] Switching the setting off has a limit:

"Your data will still be included in model training runs that are already in progress, or in models that have been trained."

— Anthropic Privacy Center, How long do you store my data? (July 2026).[16]

Team and Enterprise plans have no model training on your content by default.[2]

Gemini. With Keep Activity on, Google uses your chats to improve its services, including training generative models. A subset is read by human reviewers, and reviewed chats are kept for up to three years, disconnected from your account.[17] Activity auto-deletes after 18 months by default. With Keep Activity off, future chats are not used for training unless you send feedback, but are still kept for 72 hours.[17] Google's own notice is blunt:

"Please don’t enter confidential information that you wouldn’t want a reviewer to see…"

— Google, Gemini Apps Privacy Notice (October 2026).[17]

Work and school accounts are different. Google says Workspace content is not human-reviewed or used to train models outside your domain without permission.[18]

Tools and integrations

ChatGPT lists plugins on every plan and add-ins for Excel, Word, PowerPoint and Google Sheets, limited on Free and Go. Plus and Pro can connect to internal tools, and Business adds shared company knowledge.[1][8] For spreadsheet work specifically, see our ChatGPT vs. Gemini comparison for spreadsheets.

Claude includes connectors on every plan, including Free. Pro and above add Claude in Chrome and Claude for Microsoft 365 and Outlook, and Anthropic lists a Google Workspace add-on.[2] Team adds enterprise search across an organisation's connected tools.[2]

Gemini is strongest where your work already sits in Google. Paid plans add Gemini to Gmail, and from AI Pro to Docs, and connected apps reach Google services plus third-party tools, including Model Context Protocol (MCP) servers. Google warns it "does not monitor or secure" custom third-party connections.[17][3]

What independent evidence says

Independent evidence is thin and goes stale fast. Two sources are worth knowing.

Arena (formerly LMArena). Arena ranks models by anonymous head-to-head votes; users learn which models answered only after voting.[19] On the text leaderboard dated 15 October 2026 (8.8 million votes, 416 models, default view), Google's Gemini 4 Argon (High) was first, marked "preliminary". Anthropic models held six of the top ten places, and OpenAI's highest entry was 21st.[20] But Arena scores named models at stated effort settings, through the API. That may not be what your app or plan serves. Votes measure preference, which includes style, not correctness. And a 2025 study found that undisclosed private testing let some providers pick which scores to publish.[21] Our benchmarks explainer covers these limits.

EBU and BBC, October 2025. Journalists from 22 public broadcasters in 18 countries rated responses to news questions from the free versions of ChatGPT, Copilot, Gemini and Perplexity, generated in late May and early June 2025.[22][23] Overall, 45% of answers had at least one significant issue. Gemini had significant issues in 76% of responses, against 36% for ChatGPT, mostly because of sourcing.[23] Claude was not tested. The study is more than a year old and covered the free models of the time: it shows sourcing failures were common, not how today's products compare. Perplexity, the fourth tool in that study, is a search-first product rather than a general assistant; our Perplexity review covers how it cites sources.

AI Chronicle

Documented limits and risks

• Limits move. All three cap usage within a plan, and reserve the right to change plans and caps. Anthropic says prices "are subject to change"; Google says limits "may change without notice".[2][4] OpenAI cut the included usage for new Pro 200 subscriptions in September 2026 while keeping the price.[7]
• Inconsistent pages. Some vendor pages disagree with each other. Google's subscriptions page promotes Gemini 4 Argon, while its FAQ on the same page still names "3.1 Pro".[3] Anthropic's pricing table says "Opt-out" for consumer training, while its privacy centre describes a choice to allow.[2][15]
• Accuracy and actions. Google warns that Gemini Apps "may produce inaccurate or offensive information" and advises users to "supervise Gemini's web browsing and tasks closely".[17] Our explainer on AI hallucinations covers why confident answers fail.

Who each suits, and who should skip it

Reasoned from documentation, not performance.

ChatGPT

Suits: people who want the widest feature set in one app, including image creation, voice with video and coding agents. Also Microsoft 365 users, and small teams wanting a self-serve workspace from two seats.[1][8] Skip, or check first: heavy readers on a budget. ChatGPT Free's 27K Instant window is the smallest documented free-tier window of the three, and Free allows three uploads a day.[1][11]

Claude

Suits: people working with long documents or many files on modest plans, and teams that want connectors on every tier.[2][12] Skip, or check first: organisations that need a fixed all-in Enterprise price, since usage is billed on top of seats.[2] Also free users who want Opus or Research mode, which start at Pro.[2]

Gemini

Suits: people and companies already in Google Workspace. The $4.99 AI Plus plan is the cheapest paid entry point of the three.[3][9] Skip, or check first: anyone pasting sensitive material into a personal account with Keep Activity on, given human review and three-year retention of reviewed chats.[17] Also free-tier users with long documents, given the 32K window.[4]

How to test it yourself

A 7-step test before you subscribe

1. Pick 10 real tasks from last month, with sensitive details removed.
2. Write the right answer first, or what "good" looks like, before you run anything.
3. Use the plan you would pay for. Free-tier models and context windows differ from paid ones.
4. Set privacy first. Turn off training, or use a temporary or incognito chat, before uploading anything.
5. Run the same prompt in all three, and record the model name and date shown.
6. Check every fact and citation against the source. Note the time you spend fixing each answer.
7. Try one integration you would really use, such as a Drive file, a spreadsheet or an email draft, and note any setup and permission prompts.

Our guide on how to judge a new AI model release explains how to keep a small test fair.

A worked example (hypothetical)

This scenario is invented to show how the documented facts narrow a choice. No product was tested and no results are reported.

Amira runs a three-person bookkeeping practice on Google Workspace Business Standard. She wants help with client emails and long tax guidance PDFs.

Data first. Client data rules out personal plans unless training is off, so she looks at business tiers. At list prices on 16 October 2026, three seats billed yearly cost $60 a month on ChatGPT Business or Claude Team Standard.[8][2] Her existing Workspace plan already includes the Gemini app with expanded access.[9]

Then fit. Gemini works inside the Gmail and Docs she already uses, though she would confirm her Workspace edition's file limits, which Google documents separately from personal accounts.[13] Claude's 1,000-page PDF limit suits her longest guidance documents.[12]

Then test. She runs 10 past tasks through Gemini in Workspace and through a month of Claude Team on the two-seat minimum, scoring each answer against her own figures. Only then does she decide whether a second subscription earns its cost.

Questions readers commonly ask

Which is the best AI assistant?

We don't name one. We haven't tested them, and the documented differences favour different users. Independent evidence is limited: Arena measures preference for models through the API, and the EBU/BBC study did not include Claude.[20][23]

Are the free versions good enough?

For light use, possibly. But free tiers have smaller limits: ChatGPT Free's Instant context is 27K tokens and Gemini's free window is 32K.[1][4] Claude's pricing lists up to 1M "varies by model" on Free, but Opus is paid-only.[2]

Can I stop my chats being used for training?

Yes, on all three. ChatGPT: turn off "Improve the model for everyone". Claude: the model-improvement setting in Privacy settings. Gemini: turn off Keep Activity, or use temporary chats.[14][15][17] Business plans do not train on your content by default.[8][2][18]

Will these prices still be right next month?

Probably not all of them. OpenAI changed its Pro tiers in September 2026.[7] Check the pricing pages in our sources before buying.

The next step

Plan sheets change every few months, and so do the models behind them. Before you switch assistants after a launch, use our checklist for how to judge a new AI model release.

Sources and evidence note

Plans, prices, limits and data terms come from each vendor's pricing, help and privacy pages. They describe documented features and terms, not tested performance. Prices are US list prices before tax as displayed on 16 October 2026. Independent evidence comes from the Arena leaderboard and method page, The Leaderboard Illusion (2025) and the EBU/BBC News Integrity in AI Assistants report, each read in full on the publisher's site. The worked example is invented. All pages were accessed on 16 October 2026.

Related reading

• Perplexity vs. ChatGPT for research
• What AI benchmarks actually measure