Most “best AI writing tools” articles compare products. This one starts from three jobs, because the tool only matters once you know what you are asking it to do. Below are the three writing jobs I hand to a model most weeks, and an admission: I only hand them to one tool. That admission is the interesting part, so I have put it near the top instead of burying it.
The three jobs, and what goes in the prompt
- The first hundred words of an article. The hardest part of any post and the part that decides whether anyone reads the rest. The prompt gives the topic, the audience, one fact I want in the opening and a hard instruction: no scene-setting, no “in today’s landscape”, start on the specific.
- A short email I do not want to send. Turning down a piece of work, chasing something late, correcting a misunderstanding. Under 120 words, plain, not grovelling.
- Rewriting a paragraph I have already decided is bad. Not “improve this” — I say what is wrong with it first, then ask for a version that fixes exactly that.
If you want to compare tools on jobs like these, fix the wording of each prompt first and reuse it word for word in every tool. The twenty-minute test in my chatbot notes sets out how, and it is worth more than anything I could tell you about products I have not opened.
Who I ran them through, and who I did not
Claude, daily, mostly through Claude Code in a terminal rather than a chat window. That is the whole list. I have not subscribed to Jasper, I have not subscribed to Writesonic, and I do not use Grammarly. Those three appear further down with prices read off their own pages and nothing else, clearly labelled, because a review of a product I have never opened is not a review.
Task one: the first hundred words
This is where the model is most useful and most dangerous at the same time. Useful, because it will produce eight openings in the time I would spend agonising over one, and seeing eight bad ones makes the ninth obvious. Dangerous, because every single one of those eight will be fluent, and fluency reads like quality until you check whether the first sentence contains a fact.
What I get back, reliably, when I do not fight it: a sentence establishing that the topic is important, a sentence establishing that it is confusing, and a promise that the article will clear it up. Three sentences, zero information. The fix that works for me is putting one concrete thing in the prompt — a number, a date, a thing that broke — and forbidding any sentence before it.
Net verdict on task one: genuinely faster, but only because I throw most of it away. If you keep the first output, you have saved ten minutes and published something nobody will finish.
Task two: the email you do not want to write
I am not going to paste real correspondence into a blog post, so you get the shape of the result rather than a transcript. Take that for what it is worth, which is less than a screenshot would be.
The characteristic failures are consistent enough to list. Drafts come back too long, because politeness gets expressed as volume. They open with a sentence about hoping the recipient is well. They soften the actual message until it is no longer in there — an email declining work comes back sounding like an email postponing work, which is a different email with different consequences. And the one to watch hardest: a friendly closing line that offers something I never agreed to, a free extra or a date I cannot meet.
That last one is the reason this task stays on a short leash. A fabricated sentence in an article is embarrassing. A fabricated sentence in an email is a promise someone else now believes you made. My rule: the model drafts, I send, and I read every line of an outgoing message as if I had written it, because as far as the recipient is concerned I did.
Task three: a bad paragraph of mine, rewritten
For this one I can show you the material, because the bad paragraph is mine. Here is how the previous version of this exact page opened:
AI writing tools have moved far beyond simple grammar checkers. The current generation can draft full articles, match a specific brand voice, restructure clunky paragraphs, and adapt tone for different platforms in seconds. But not every tool is built for the same job, and picking the wrong one wastes both money and time.
Nothing in it is false. Nothing in it is information either. It could sit at the top of any article about any category of software with four words changed. I asked for a rewrite that cut the throat-clearing and led with something specific, and what came back was shorter, tighter, better rhythm — and still empty, because the input contained no facts and a rewrite cannot manufacture them.
That is the single most useful thing I learned rebuilding this site, where a hundred posts are being cut to twenty-five. These tools are strong at compression and rhythm and completely unable to add substance. The paragraph only got better when I replaced it with something that could not have been written about any other topic: three named tasks, one tool, and an admission about the other three. The model did the tightening. I had to supply the sentence that was worth tightening. The whole ugly story of what happens when you skip that step is in the post about the mistakes that got this blog rejected.
What they cost, where I could read the page
| Tool | Free option | Paid, as published | Tested by me |
|---|---|---|---|
| Claude | Free tier, $0; no fixed message count published | Pro $20/month, or $17/month billed annually at $200 up front. Max from $100/month | Yes, daily |
| Jasper | None. A “7-day free trial” on the Pro plan | Pro $59/month per seat billed annually, $69 billed monthly. Business is custom-priced | No |
| Writesonic | None. The page says “Start free. No credit card needed.” but lists no free plan | Starter $79/month, Basic $199/month, Growth $399/month, all billed annually. Enterprise custom | No |
| Grammarly | Free tier, €0, including spelling and grammar, tone, and “Generate text with 100 AI prompts” | Pro €12/month. Enterprise is contact-sales | No |
What the old version of this page got wrong
Two things, and the pricing table above is what exposed them.
The first: this page used to end with a box crowning Jasper as the number one pick for AI writing, a tool the article above it never mentioned and I have never used. The story of those boxes, with the counts, is in the affiliate mistakes I made on this blog.
The second is worse. A sister page on this site called Writesonic “our top free-friendly pick” and praised its “generous free plan”. When I opened the Writesonic pricing page on 20 September 2026, there was no free plan on it. The cheapest tier listed was $79 a month billed annually. Jasper does not have a free plan either — a seven-day trial is not a free plan, and the difference is a payment. Neither claim was tested. Neither was checked. Both were the kind of sentence that gets written when you are filling a template rather than reporting something.
I have no affiliate relationship with any of these companies, so nothing here pays me whichever way it falls. When that changes, links will say so next to the link, not in a footer.
The pass nothing can do for you
After the draft comes the part that takes the longest and never appears in tool comparisons: checking whether each claim is true, replacing the generic sentences with things only you know, cutting the paragraph that exists purely to reach a word count, and reading it aloud to find the sentences no human would say. On a post this length it is the longest part of the job. It is also the entire difference between a page worth publishing and the sort of page that gets a site turned down for low-value content.
If you are choosing a tool, choose on task one and task three — drafting speed and rewriting quality — and assume the rest of the work is yours. The method I use now for a whole article, rather than a paragraph, is in the guide to writing SEO articles with AI without publishing filler. And if you want the same treatment applied to code instead of prose, the coding assistants post covers the tool I actually use all day.