The week a model disappeared
Anthropic had to suspend access to two new models abruptly in June and restored it after a political and security dispute. Whatever one thinks of the particular case, the product lesson was remarkably clear.
A model can be excellent and tomorrow still be unavailable, disallowed, too expensive, or unsuitable for a particular data context. If the whole product is confused with a model name, there is no strategy. There is a bet.
A model is an exceptionally important component. That is exactly why it cannot be the entire architecture.
Model loyalty is not customer value
Teams develop preferences quickly. One model writes better, another works more cheaply, one feels calmer around tools. That experience is valuable. It becomes a problem when a current preference is embedded into a permanent product promise.
Customers are not buying a romantic relationship with a model provider. They are buying a job that should be done reliably. The model name may be part of the explanation, but not the only answer to reliability.
Choice means more than a dropdown
Listing several models is easy. Real choice appears when quality, cost, speed, data requirements, and availability can be weighed without reinventing the whole use case.
The goal is not to claim models are interchangeable. They are not. The goal is to treat differences as conscious product decisions instead of surprising infrastructure crises.
- The use case is described independently of a marketing name
- Quality, cost, and data needs remain separate questions
- A switch is a deliberate, tested decision
- A provider problem automatically stops the product
Local models belong to resilience too
Self-hosting is often discussed only as a privacy choice. It is also operational independence. Not every task needs the largest model, and not every organization wants critical work to depend entirely on external access.
That does not mean rejecting cloud models. It means using their extraordinary capability as a conscious service choice, not as an irreplaceable condition for the product to exist.
The product promise must outlive the leaderboard
Frontier models change faster than buying cycles, contracts, and business processes. A durable agent product is therefore built on tasks, rules, knowledge, and accountability — and uses the right model as its engine.
Love your favorite model. Just do not build an operation that cannot get up without that love.