1/ On 23 October, OpenAI switched off a dozen legacy API models in one day, GPT-4, o1 and o3-mini among them, with no automatic redirect. Another wave lands in January. Every model you depend on has a shutdown date. Here are 9 layers of a model-proof AI stack 🧵
2/ Move 1, own the dependency: keep a live inventory of every model ID, caller and owner. Make apps call a stable alias you control at the gateway. Put every shutdown date on a calendar with a named owner.
3/ Move 2, qualify the successor: test against golden evals per workload, not leaderboards. Run the new model silently on real traffic. Diff tool calls, output formats, latency and cost before anything goes live.
4/ Move 3, swap without drama: shift traffic in weighted steps, make rollback a weight change rather than a redeploy, and contract for exit with notice periods, capacity and an open-weight fallback.
5/ Case study: GitHub retired several Copilot models on 19 Oct with a month's notice, a named successor for each, and a nudge for admins to enable them. Lesson: model choice is policy, and someone must own the switch. How many model IDs are hard-coded in your stack?
