Learn · 7 of 7
The landscape, briefly
Who makes these, how they differ, and what actually constrains your choice.
A handful of labs train frontier models (The largest and most capable model a vendor currently offers, priced accordingly.); a larger number of companies serve them; and the part that actually constrains you is usually neither capability nor price.
The shape of it
Every vendor ships a family rather than a model: a flagship, something mid-sized, and something small and fast. The flagship is typically five to twenty times the price of the smallest in the same family, and on a great deal of ordinary work the small one is indistinguishable.
Model names move constantly and mean little from the outside. A version number tells you the vendor’s ordering, not whether it suits your task.
Open weights
Some models publish their trained parameters, so you can run them on your own hardware. That is not the same as open source — the training data and code usually are not published, and licences often carry conditions — which is why “open weights (A model whose trained parameters are published, so anyone can run it on their own hardware.)” is the more accurate phrase.
The reasons to want it are privacy, control, and not being subject to a vendor deprecating something you depend on. The cost is that you now operate inference (Running a trained model to get an answer. Everything you pay for by the token is inference, not training.) infrastructure, which is more work than an API key and rarely cheaper until volumes are high.
What actually constrains the choice
Rarely the model. Usually one of:
- Rate limits (A cap on how much you may send in a window, usually counted in requests and tokens per minute.). How much you may send per minute, which breaks more prototypes than quality ever does, and appears in no comparison table including this one.
- Deprecation. Models are retired. Anything you build on a specific version needs a plan for the day it goes away.
- Where the data goes. Often the decision that gets made for you, by somebody else.
How to stay current without chasing it
Most announcements do not change what you should do. The ones that do are price moves on models you actually run, deprecations with a date, and genuinely cheaper like-for-like options. That is what the log records, and it deliberately records nothing else.