API deprecation for AI models: what breaks when a model is retired

As AI vendors like OpenAI and Anthropic accelerate their release cycles, model deprecation has become a critical concern for developers. Unlike traditional API endpoints that simply change structure, retiring an AI model means the underlying service ceases to exist, leading to 404 errors or silent performance drift if using model aliases. OpenAI typically provides at least six months of notice for general models, while Anthropic maintains a 60-day window and includes tentative retirement dates at launch. The article highlights that closed-source models are entirely dependent on vendor infrastructure, whereas open-weights models can be preserved via decentralized methods like BitTorrent. Developers are advised to pin specific model snapshots, maintain robust evaluation suites, and implement multi-vendor fallback strategies to mitigate the risks of sudden service termination. Ultimately, the article emphasizes that if a model is essential to your infrastructure, you must be able to run it yourself.