In the first article we wrote that you decide what not to hand over first. With that settled, this one is about which one you actually use.
Where this article stands, before anything else
The draft of this article was written with Claude.
Which means the text you are reading contains writing produced by asking Claude to compare Claude, ChatGPT and Gemini. We are in no position to rank them fairly.
So this article does not rank them. It does not say which is cleverest, either. All it sets out is how we actually use each of them. The judgement is yours.
This is not limited to this article: at the bottom of every OS Notes piece we state how far AI was involved.
Why there is no comparison table
We started making one and stopped. Two reasons.
It dates immediately. These three release something new every few months. Put what we found on the day we checked into a table, and readers end up deciding on old information.
The way we use it is not the way you use it. We lean heavily on “read something long” and “build the app”. For a company working with drawings, or one handling large numbers of photos, the answer will naturally be different.
Instead, at the end we set out how to check it for your own company. That should be more use.
How we split it up
Not a ranking — just which one we open, and when.
When something long needs reading
When what needs reading is long and turns on the detail — specifications, contracts — we open Claude. It handles a long document all the way to the end, and in our experience drops less of the middle.
Same when writing the app. That is how our business app gets built.
When we just want to ask something
For asking whatever comes to mind on the spot, it is usually ChatGPT. More people use it than anything else, which makes it easier to say “use this” to someone in the company. That is the real advantage.
That is not about performance, it is about the effort of teaching it. Whether a tool you hand to someone sticks depends on whether the people around them use it too. Same as the tools on site.
When the work is already inside Google
The spreadsheet behind an estimate, a long thread that arrived in Gmail, a document on the shared drive. When what we are working on is already inside Google, we open Gemini.
Not having to paste the file across each time is the big thing. This is a case where where the file lives matters more than the difference in performance.
The ones that have just come out
New ones keep arriving. Trying them is a good thing in itself.
But the cost of switching is not the price. It is the time spent relearning how to ask so that it does what you meant. That is the expensive part.
So the rule we settled on is: try new things when they come out, but do not move real work over lightly. It is like changing your tools — the speed you have with the one you know is nothing to sniff at.
Questions we get asked
“Can we not just stay on the free plan?”
While you are trying it out, it is plenty. It is different once company information goes in. Personal plans, free or paid, and business plans treat what you put in differently. The third article covers all of that.
“Are they not weak in Japanese?”
All the main ones handle Japanese perfectly well now. Where the difference shows is trade language.
“Genjo-kaifuku” (restoration), “who pays for what”, “shell”, “plasterboard substrate”. Used as they are, terms like these are sometimes taken in their everyday sense. Set out the background first, then ask, and the accuracy changes.
For example, we write this.
You are a member of staff at a contractor that handles restoration of rental properties in Japan. “Restoration” here means the works that return a unit, after the tenant has moved out, to a state in which it can be let to the next one. Read the email below and list, as bullet points, only the items that we are responsible for paying for.
This is no different from a works instruction. Hand it over with no background and a “sort it out, please”, and what comes back is not what you had in mind.
“Should we settle on just one?”
If it is going to be used across the company, we would start with one.
With several, you get “I cannot remember which one I asked” and “the way I phrased it there does not work here”. The smaller the company, the better it sticks when there is less to remember.
We use several because we also use them for development. For ordinary office work, one is enough.
How to check it for your own company
Faster than reading comparison articles, we think.
1. Take one real document of your own
A long email that actually arrived from a property manager, or one of your own specifications. Take out personal names, property names and addresses.
2. Give every candidate the same task
Phrase it so that the answer can be marked — something like “list, as bullet points, only the items in this document that we are responsible for paying for”. Asking for opinions gives you nothing to compare.
3. Check against the original and count the mistakes
Things missed, things added that were never there, things misread. Count them and the ranking appears by itself. And it is your company’s ranking, not somebody else’s.
4. Use it on real work for a week
It comes down to whether you feel like opening it, more than how clever it is. Something you stop using is worth nothing, however clever it is.
The next article is the important one. It is about what happens to the information you put in, and the shape of the pricing.
This series
- Decide what not to hand over, before you pick one
- Using Claude, ChatGPT and Gemini for different jobs (this article)
- What happens to your information, and what it costs — what to check before you sign