The cost of leaving an API belongs in the decision before its shutdown notice arrives. Useful features on day one do not pay for rebuilding an integration. When the calendar forces a switch, whoever depends on it still has ongoing work to deliver.
The Assistants API shut down on August 26, 2026. OpenAI directed developers to Responses and Conversations. That deadline deserves a place beside the capability list: a dependency can still serve a product while its provider forces developers to change how they use it.
After that week, on September 15, 2026, I discussed the friction of my move from Claude to Codex. I considered the Codex model better, while Claude felt ahead in that practical experience. Switching applications differs from migrating an API. What the experience makes me demand is attention to the work required to rebuild a workflow that was already functioning.
An integration accumulates decisions outside the model selector. How is a session identified? Where does an interrupted execution resume? What needs to survive from a conversation to serve the user? If this information only makes sense inside the provider's interface, leaving first requires discovering the meaning of what was stored there.
I prefer separating product concerns from API conveniences early. A task's objective and artifacts need to remain understandable. An enormous abstraction written before any real work exists also charges its own maintenance cost. I would start with the data that actually needs to survive the switch and a way to retrieve it in a useful form.
Getting the migration to compile settles only part of the verification. An agent can keep answering while changing how it works. To check that, I want representative tasks with expected results recorded before the change. Wording can vary; the behavior satisfying the request needs to remain verifiable. That is where a transition becomes more than an adjustment to API calls.
An announced deadline helps reserve time, but a reminder performs none of the migration. When assessing a dependency, I want an understandable route for carrying an ongoing task across: its objective, artifacts, and enough information to continue. I will include that effort in the integration cost. If leaving requires explaining everything from scratch, the entry price was concealing a substantial part of the work, and a better model at the destination will not do that reconstruction for free.