· 1 min read
Models shipped faster than the product
In 2026 the model stopped being the stable part. The product around it still has to move on purpose.
The useful change this year was not one model. It was the cadence. A default from the spring was a fallback by autumn. The product around it did not move at that speed. Prompts, tools, evals, and the one screen a person actually uses all lag the release notes.
What actually broke
A silent model upgrade looks like a good week until a path you trusted starts hedging, skipping a tool, or answering a question the screen never asked. Nothing in your diff changed. The dependency did.
I started treating the model the way I treat any other library I do not control.
- Pin it where you pin dependencies. A floating alias is an outage with better marketing.
- Keep a small eval that fails when the product breaks, not when a public benchmark moves.
- Let the demo chase the new model. Change the system when that eval says so.
The demo and the system are different artifacts. Mixing them is how a Tuesday release becomes a support queue.
What I stopped doing
I stopped rewriting the product every time a lab posted a chart. Most of those charts measure a task I do not have. The tasks I do have are narrower: call this tool, refuse that input, return a shape the client already parses, stop when the budget is gone.
A newer model is not a migration. A migration has a diff, a rollback, and a reason.
The frame stays empty on purpose. You fill it when you mean to, not when the next card lands on the pile.