Your model choice is the least defensible thing about you.

There is a specific quarter you remember. You switched to the newest model the week it shipped, ran your usual work through it, and felt something real. The drafts came back sharper. The reasoning held together for longer stretches. For about six weeks you had an edge. Then your competitor switched to the same model, and the edge became the floor everyone stands on.

That is the treadmill, and there is now a clock on it. Epoch AI measures how far the best open-weight models trail the closed frontier, and through the first half of 2026 the gap sits at roughly four months. It moves like a sawtooth. Open nearly draws level, a closed lab ships something, the gap reopens, open closes it again. Four months is not a moat. It is a delivery date.

Capability is the input that converges fastest

Every frontier model is a rental. You do not own it, you did not train it, you cannot keep the version you liked when the lab retires it. Anyone with a credit card rents the same one, from the same short list of labs, at a price printed on a public page. Whatever you can do with it this quarter, the person you are competing against can do next quarter, and the labs are spending billions to guarantee it. The thing everyone shops for is the thing that levels out across everyone at once.

Which makes the model you pick the least defensible thing about you. I know that reads as wrong, because model choice is where nearly all the attention goes: the benchmark threads, the quiet pride in running the newest thing the week it lands. Spend an afternoon reading how people talk about their setups and you would think the model was the whole game. It is the most expensive way to stay exactly as replaceable as the next person, because you are pouring judgment into the one variable guaranteed to be a commodity by autumn.

One thing on your side does not converge

It is the accumulated context the model runs on. The decisions you made and the ones you reversed, the corrections that finally stuck, your house style, the thousand small judgments about how the work gets done in your shop and not in the abstract. None of that arrives in a lab’s next release. It cannot be rented, because nobody else has it. Point two different models at a rich enough record of how you actually work and they start to converge on your answers rather than on each other’s.

Almost nobody invests there. The record of how you work gets treated as exhaust, scattered across chat histories that expire and prompts retyped, from memory, into every fresh session as though the last one never happened. Meanwhile the interchangeable input gets treated as the asset worth agonising over. That is investing backwards. You are guarding the thing that levels and discarding the thing that compounds.

The asymmetry is the argument. Model capability plateaus across the whole field on a schedule you can read off a chart. A corpus of how your work gets done compounds for exactly one party. Every correction makes the next output a little better, and the improvement stays with you when the model underneath gets swapped for a cheaper, faster one next year, which it will.

The part worth defending should be the part you can reach

This is why the thing we are most stubborn about in Brief is not which model sits behind a given associate. That is a setting. It should be visible, and you should be able to change it whenever a better one shows up. The stubborn part is that the context your staff accumulates is yours, held in an open format, exportable whole on any ordinary day. Not as a courtesy. Because it is the only thing in the whole arrangement that was ever going to be worth defending, and a thing worth defending should not live somewhere you cannot get it back.

So pick whatever model tops the benchmark this month. It is a good model. Next month there will be a better one you can rent for less. Just stop mistaking the rental for the asset, and put your attention on the record only you can build. Let the model be the part you replace without a second thought.