I have read a lot of chatbot comparisons that claim months of side-by-side testing, and I used to publish one of them. The earlier version of this page opened with the line that no single chatbot wins more than half the categories “in blind task-by-task comparisons”, and then handed out verdicts for writing, coding, research and document analysis. I never ran those comparisons. Nobody here did. That page is gone, and this is what replaces it.
What this page is, and what it is not
It is not a head-to-head test. I use one of these four every working day and have not tested the other three, and I won’t pretend this is a lab. What it is: a decision guide built from two things I can stand behind. First, what each vendor publishes about its own product, read off the page on a stated date. Second, my own experience with the one tool I genuinely live in, described concretely enough that you can judge whether my work resembles yours.
If you want a winner declared, you will be disappointed. If you want to stop paying for three subscriptions you barely open, keep reading.
The one I actually use, and what I use it for
Claude is the only one of the four that is open on my machine most days, and almost always through Claude Code in a terminal rather than the chat window. The work it does for me is unglamorous and very specific: reading a folder of exported WordPress posts and telling me which ones are duplicates of each other, writing a shell script that measures response times across my own sites, rewording a paragraph I already know is weak, and editing WordPress in bulk through the REST API instead of clicking through ninety admin screens.
The honest shape of that relationship: it is fast at the mechanical half of the job and useless at the half that matters. It will restructure a post in seconds. It will also, left alone, produce exactly the kind of well-formed, confident, empty writing that got this site rejected by AdSense four times. I have written up that mess in detail in a full count of those mistakes, because it is the most useful thing on the site.
Where it fails, consistently, in my own use: it invents plausible detail when it does not know something, it agrees with me far too readily when I push back on a correct answer, and it will happily write “in my testing” into a draft if I do not stop it. That last one is not a quirk. It is the single reason this site needed rebuilding.
What the pricing pages said when I opened them
I opened each vendor’s own pricing page on 20 September 2026 from a home connection in Spain, and copied down what was on it. Two of the four let me read the page. Two did not.
| Plan | What the page said | Source, checked 20 September 2026 |
|---|---|---|
| Claude Free | $0. No fixed message count published; the page describes it as covering “everyday questions” | claude.com/pricing |
| Claude Pro | $20/month, or $17/month billed annually at $200 up front. Described as “at least 5x more usage per 5-hour session than Free”, with weekly limits on top | claude.com/pricing |
| Claude Max | From $100/month, in two tiers described as “5x or 20x more usage than Pro” per 5-hour session | claude.com/pricing |
| Google One, free tier | 15 GB storage, no paid Gemini allowance | one.google.com/about/plans |
| Google AI Plus | €9.99/month, 2 TB, Gemini app limits “2x higher than without a Google AI plan”, Flash Thinking model, Gemini in Gmail and Notebook | one.google.com/about/plans |
| Google AI Pro | €21.99/month, 5 TB, limits “4x higher”, Pro model, Deep Research, Gemini across Gmail, Docs, Vids and Notebook | one.google.com/about/plans |
| ChatGPT | Not quoted. The pricing page returned HTTP 403 to my request on the day | openai.com/chatgpt/pricing |
| Perplexity | Not quoted. The pricing page returned HTTP 403 to my request on the day | perplexity.ai/pricing |
Three things in that table are worth more than the numbers themselves.
- The “about 20 USD monthly” shorthand everyone repeats is already wrong in two directions. Claude Pro is $20 monthly but $17 on the annual plan, and the annual plan wants $200 in one payment. Google’s nearest equivalents were €9.99 and €21.99 when I looked, which is not one price, it is two, and neither is twenty dollars.
- Google sells storage with a model allowance attached, not a model subscription. Both paid tiers I could read lead with 2 TB and 5 TB. If you already pay Google for storage, the marginal cost of the AI is much smaller than a standalone comparison suggests. If you do not, you are buying storage you may not want.
- Limits are published as multipliers, not as numbers. “5x more usage than Free” and “2x higher than without a Google AI plan” are the vendors’ own words. Neither tells you how many messages you get. Anyone who prints a hard message limit in a comparison table either has inside information or made it up.
Two of them would not let me read the page at all
OpenAI’s and Perplexity’s pricing pages both refused my request with a 403 on the day I checked. That is not a scandal; large sites block automated requests routinely, and both pages open fine in a normal browser. But it changes what I am allowed to write. I could open them by hand and type the numbers in, and if I were only covering two products I would. Across a page like this one, the number I typed six weeks ago is the number that quietly goes stale and turns a useful article into a misleading one.
So: no ChatGPT price on this page, and no Perplexity price. Two links instead, and a date. That is less satisfying than a full table and considerably more honest than the version of this page that used to sit here, which gave all four a confident “about 20 USD monthly” without a source or a date attached to any of them.
Which one for which job, and how I know
Here is the table I can defend. The right-hand column is the part most comparisons leave out, and it is the only column I would actually read if someone else had written this.
| Job | Where I would start | How I know that |
|---|---|---|
| Editing code across several files in a real project | Claude, through Claude Code | Daily use on my own WordPress blogs and a small Node service I run on a VPS |
| Rewriting text you have already judged to be weak | Claude | Daily use, including every post in this rebuild |
| Anything that has to be current this week | A tool that cites live sources, so Gemini or Perplexity | Not tested by me. Based on what the vendors publish about search grounding |
| Drafting inside Gmail and Docs | Gemini, if you already pay Google for storage | Not tested by me. Based on the plan contents listed on Google’s own pricing page |
| A first draft of something short and disposable | Any of the four free tiers | Opinion, not a measurement. They are all far past the quality bar for a first draft |
| Anything you will publish under your own name | None of them, unaided | Four AdSense rejections on this site |
Notice how much of that column says “not tested by me”. That is the true state of my knowledge, and I would rather show it than pad the table.
The twenty-minute test that beats this article
Every comparison you read, including this one, was written by someone whose work is not your work. The fix is embarrassingly cheap: all four have a free tier, so build your own benchmark once and rerun it when something big ships.
- Pick four tasks you actually did last week. Not clever prompts. The boring ones. Mine would be: tidy a messy paragraph, explain an error message from a web server, turn a list of post titles into groups by topic, and find the source for a claim I half-remember.
- Write each one down as a prompt, once, and save the file. The value is in running the identical prompt across tools and across months. Improvising a new prompt for each tool is how people convince themselves of things.
- Run all four, paste the outputs into one document, and leave it overnight. Reading them cold the next morning is startlingly different from reading them as they stream out.
- Count edits, not impressions. How many sentences would you have to change before you would put your name on it? That number is the whole result.
- Do it again in six months. Not six weeks. Model releases are constant and switching costs you your saved prompts and your habits.
Three ways this kind of comparison goes wrong
Having written a bad one, I know the failure modes from the inside.
- The verdict outlives the evidence. A page written against one model version keeps ranking after two more have shipped. Every date on this page exists so you can see how stale it has become.
- The prose smuggles in a test that never happened. “In our testing” and “after months of running the same tasks” are four words that cost nothing to type and completely change what a reader believes. They were all over the previous version of this page.
- A sponsor sets the verdict. The old version of this page ended with a promotional box naming ChatGPT the site’s number-one choice, with an affiliate-style label attached, while the body text said no chatbot won more than half the categories. The box and the article disagreed, and the box was the part that was designed to be clicked. There is no such box now, and I am not in any affiliate programme at the moment, so nothing on this page pays me if you click it.
What I would actually do
Start free, all four, for a fortnight, on the tasks you already have. Then hold one paid plan, chosen by where your own edit count dropped the most, and check the price on the vendor’s page rather than in an article. If you already pay Google for storage, look at what your existing plan includes before you add anything at all.
And treat the output as a draft with no authority behind it. The part that makes a page worth reading is the part a model cannot supply: what you measured, what you got wrong, what you would not do again. That is also, as far as I can tell after four rejections, the part search engines and ad reviewers are looking for. I go through how I now write with these tools without producing filler in the piece on writing SEO articles with AI, and the tool-by-tool detail on Claude itself is in the longer Claude review. If your interest is specifically code rather than prose, the post on AI coding assistants is the more useful one.