Four differences, and only one is the model
Assistants differ in the model, the hidden instructions, the filter and the data policy. The model is the part everybody names and the part that varies least between serious products; the other 3 explain most of what people notice day to day.
Comparisons usually start and end with which model. That is the hardest thing to evaluate and the least likely to explain why one product feels better than another for your work.
The model is the smallest difference in practice
For ordinary text work, the leading models are close enough that most people cannot tell them apart in a blind comparison. Differences show at the extremes: very long reasoning, unusual languages, specialised knowledge. If your work is not at an extreme, this is not where your answer lies.
The hidden instructions are the largest
A block of text placed before your first message sets caution, length and manner. It is invisible, it is different in every product, and it changes the feel of the same model completely. This is why 2 products on identical models are described as having different personalities.
The filter decides what you can ask
Independent of the model and set by the operator. A permissive product on a cautious model and a cautious product on a permissive model both exist. Judging the filter from the model, or the reverse, is the most common mistake in a comparison.
The data policy decides what happens afterwards
Retention, training and account requirements vary far more between products than the models do, and they are written down rather than inferred. This makes the data policy both the most consequential difference and the easiest difference to check.
What people ask before switching
Should I just pick by model then?
Only if your work sits at an extreme. Otherwise the instructions, the filter and the policy will decide your experience.
How do I see the instructions?
You cannot read them and you can measure them: ask for a 1-sentence answer and see whether the length instruction beats yours.
Do products change models quietly?
Regularly, and rarely with an announcement. This is another reason to compare behaviour rather than model names.
Compare the 3 things that are not the model.
Open the chat