An Australian technology adviser has argued that a country gets AI independence by keeping the freedom to choose and replace models, not by building a national one, and that small models for ordinary jobs are the practical way to keep that freedom. The argument is about governments, but it holds just as well for a shop, a clinic or a logistics firm.
Most switches are not a dispute
Windsurf's cut-off made news because the company was well known. Most changes are quieter and more routine. OpenAI emailed its API customers that it would end access to the chatgpt-4o-latest model on February 16, 2026, about three months away, and every application built on it had to move to a newer model. Nobody had quarrelled with anybody. A model simply reached the end of its life.
That is the useful lesson. Models are replaced on the provider's schedule, so any business that uses one will switch several times, whether or not there is ever a fight. The only open question is whether a switch is a week of panic or an afternoon of checking.
What makes a switch cheap
A switch is painful for one reason: nobody can tell whether the new model does the job as well as the old one. The instructions were tuned to the old model's habits, and nobody kept a record of what a right answer looks like.
For example, a regional distributor might use AI to sort 300 supplier emails a day into four piles: invoices, delivery delays, price changes and complaints. If someone keeps 200 real emails with the correct pile written next to each, then a replacement model can be tested on all 200 in an hour. If it gets nearly as many right as the old one, the distributor switches. If not, it tries another. I would estimate that turns a switch from a project of several weeks into a day's work, and the list of 200 emails is the only thing that had to be built. It is the same point as the advantage coming from your own data: the model is the part you can replace, and the examples are the part only you have.
Why small models make this easier
Sorting emails does not need a model that can also write poetry and prove theorems. A small model, one tuned to a narrow job, can do it on a modest server that the business controls, or on one rented inside its own country. Nobody can withdraw it, and if a better small model appears, trying it is cheap. Small models still need someone to run them, and they will not handle the hard cases, such as an unusual contract clause or a long research question. Those can go to a large model from one of the big providers, used on purpose and not by default.
Alternatives are also closer than they were. In March 2026 the best American model led the best Chinese one on a popular head-to-head ranking by only 2.7%, so a replacement worth testing nearly always exists. Most companies already use more than one model, and the ones that do it deliberately, with a small model for the routine work and a large one for the rare hard case, can leave a provider more easily.
The same applies to a council or a hospital group, only with higher stakes. The sensitive, repetitive work is where a small model they control is the right fit, and it is also the work they can least afford to lose overnight.
Collect the 200 examples for one job before you need them. After that, any model from any company is something you can try in an afternoon, and a change in the provider's terms stops being an emergency.