Your AI Model Has an Expiry Date: How to Design AI Systems You Can Swap

On 14 October, OpenAI announced that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini will leave its API on 14 January 2027. Fine-tuned versions of those models go with them. That is 92 days' notice, while the same OpenAI page states a minimum of six months for generally available models.

A week later, on 21 October, a compromised release of LiteLLM, a widely used open-source AI gateway, went live on PyPI. Version 1.107.0 carried a credential stealer and a remote code loader. It was up for about an hour and forty minutes before it was removed, and LiteLLM's advice to anyone who installed it was to rotate every key the host could reach.

Two dependencies you don't control

These are different problems with the same lesson. Most AI products sit on a stack the company doesn't own: a model version that has an end date, a provider API, and often a gateway or SDK in between. Each of those layers runs on someone else's schedule. The only parts you control are your product and the interface between your product and the model.

Teams that built straight against one provider's SDK, with the model name hard-coded in a dozen places and no way to test a replacement, now have a three-month project they didn't plan for. Teams that installed their gateway without pinning a version had to assume their keys were exposed.

Five rules for AI you can swap

• Keep a model inventory. Every feature, the exact model version it calls, and any fine-tunes. You can't plan a migration you can't see.
• Call models through one seam. Route every call through your own interface, so changing a model is a configuration change, not a code search.
• Own your evals. Keep a test set from your real use, so you can prove a replacement model works before you switch, rather than hoping.
• Pin and verify the gateway. Install exact versions, review upgrades and avoid pulling "latest" into production.
• Scope and rotate keys. Each key should reach only what it needs and be replaceable in minutes, not days.

The swap test

If your model disappeared on 14 January, could you list every feature that depends on it, switch to another model with a configuration change, prove the replacement works with your own tests, and rotate every key your gateway can reach today? Any "no" is your first fix. If you run these models through Azure, check Microsoft's own retirement schedule, which may differ.

How we approach it at Zylen

Zylen is an AI optimization partner. We work inside the AI systems companies already run and improve cost, security and data control one measured change at a time. Our ten-day AI Optimization Audit maps the models and keys your AI depends on and shows you what to fix first.

Design for the swap: start the AI Optimization Audit →

Sources: OpenAI API deprecations page (14 Oct 2026); LiteLLM security incident post (21 Oct 2026).