The Model Your Product Runs On Has an Expiration Date

Your product depends on something a vendor can switch off, and they have already published the date.

I mean the AI model. Azure OpenAI publishes a retirement date for every model version, and OpenAI and Anthropic retire models routinely. The one you launched on will be gone or repriced within two years, whatever you do. I’ve told founders not to build for a cloud move that never comes. This one is on the vendor’s calendar.

What you do control is how much of your code knows the model’s name.

On GoFixly, an AI office-automation product where I was hands-on fractional CTO for 13 months, every model the product can use is a routing-table entry in config: the name, the Azure OpenAI deployment, and the per-model price. Swapping models is a config change, and because the price sits next to the name, I know what the swap costs before I make it. On a recent AI product build, routing is config-driven across Azure OpenAI and Azure AI Foundry, the model’s output is checked against a schema, and 1,550+ automated tests run on every change. A swap there gets verified by a test run instead of by hoping.

In the AI-built codebases I review, from Lovable, Replit, Bolt, and Cursor, the model name is usually hardcoded next to the API key, in several places, and nothing checks the shape of what comes back. Those founders find out what a swap costs when the deprecation email arrives.

Ask your team what happens to the product when this model is retired. “A config change and a test run” is the answer you want. A long pause means you have a project nobody budgeted for, with a deadline the vendor picked.

Do you know which model your product runs on today, and when it retires?

#FractionalCTO #StartupTech #SoftwareArchitecture #StartupFounder