How to choose an assistant

Why there are no product names here#

There are no product names on this page, and that is not squeamishness. A tool recommendation is a product name attached to a price and a feature list, and this site has one method for anything of that shape: print it on Model facts, cite the vendor's own page, and diff it against that page every morning. There is no equivalent source of truth for “which assistant is best”. It would be the one data-bearing thing here with nothing to check it against, which is exactly the kind of claim this site exists to distrust.

Recommendations also rot faster than anything else. The free tiers move, the limits move, and a page confidently naming a winner is wrong within months while still reading as advice. So this page teaches the choosing instead. The criteria below are mechanical: each one is a thing you can check yourself in a few minutes, and each is here because a mechanism explained elsewhere on this site makes it matter.

Six things you can check yourself#

• Can it search, and does it say when it did?

  A model on its own is a snapshot. An assistant that can search is a different proposition from one that cannot, and the difference is invisible unless the interface tells you which just happened. Both halves matter: the ability, and the disclosure. The mechanism is on Tool use, and why a snapshot goes stale is on Training cutoff.

• Can you see what it actually did?

  When a tool runs, can you see which one, with what, and what came back? An assistant that shows its working here is one you can audit; one that silently summarises is one you have to trust. See what can go wrong, from the outside.

• Does the vendor publish a knowledge cutoff?

  Not whether the assistant will tell you — it does not reliably know — but whether the company publishes one you can look up. A vendor that publishes the date is telling you something checkable about the product. One that does not has decided you do not need to know. See what to do about it.

• Does it ask before it sends, spends, deletes or publishes?

  The moment an assistant can act rather than only write, the question stops being about answer quality. Find out what it will do without asking, and whether that is configurable, before you connect it to anything that matters. This is the single criterion worth failing a tool over. See the risk that comes with the power.

• What does it cost once you outgrow the free tier?

  Nearly all of them are free to start, and the free tier is chosen to be enough to get attached to. The question is what the next step up costs and what it lifts — usage, limits, or features you have already started relying on. Look this up on the vendor's own pricing page rather than in a review; it is the figure most often quoted stale.

• What happens to what you type?

  Retention, training, and deletion, which are three separate questions and are answered separately. They have their own page here: what happens to what you type.

How to spend the week#

Start here suggests using one deliberately for a week, which is better advice than any comparison, and this is how to make the week count. Pick one real task you already do often — not a test question, and not something you would never actually need. Do it with the assistant instead of the way you normally would, every time it comes up.

Then judge it on the criteria above rather than on how impressive the first answer was. First answers are uniformly impressive; that is what these systems are best at. What separates them is what happens on the fourth day, when you need to know whether it searched, what it did, and whether it is about to send something on your behalf.

One thing worth doing in that week: give it a task you already know the answer to, and see how it fails. Every assistant fails, and finding out how yours does it while the stakes are zero is worth more than any review. The page on checking an answer is the method.