The Five Layers of AI Vendor Lock-In

There is a comforting narrative in enterprise AI: "We can always switch models." The API is standardized. The prompts are just text. If a better model comes along — or a cheaper one, or one with better compliance properties — you point your application at the new endpoint and move on. This narrative is dangerously false.

The warning signs are already visible. OpenAI retired GPT-4o from ChatGPT on February 13, 2026, while explicitly leaving API access unchanged — a reminder that product surfaces and API surfaces have different lifecycle risks. More importantly for production teams, OpenAI deprecated the Assistants API after reaching feature parity in the Responses API, with shutdown scheduled for August 26, 2026.

This whitepaper maps the five distinct layers where AI vendor lock-in occurs — API integration, prompt engineering, fine-tuned models, embedding models, and the operational control plane — explains what breaks at each layer, quantifies the switching costs, and presents the architectural patterns that make model portability an engineering discipline rather than a fantasy. It includes a real case study from our own GPT-4o to GPT-4.1 production migration.

What You'll Learn

• The five-layer model of AI vendor lock-in, from shallow API differences to deep embedding incompatibility
• Why a baseline prompt optimized for GPT-4o scored 4 percentage points worse on GPT-4.1 — and how evaluation-driven migration recovered an 11.5-point improvement
• Why switching embedding models means re-indexing your entire corpus (potentially terabytes of vector data) with no clean shortcut
• The LLM Gateway pattern and what it solves versus what it doesn't — gateways fix Layer 1 but leave Layers 2-5 untouched
• A concrete migration playbook: audit, evaluate, re-optimize, parallel deploy, document

Who This Is For: CTOs, platform engineers, and technical leaders managing production AI systems who need to plan for model transitions without breaking things.