LLM Vendor Lock-in: How OpenAI and Anthropic Trap Enterprise Customers

OpenAI retired 33 models on January 4, 2024, including GPT-3 and all fine-tuned models based on deprecated bases, causing disruption for teams relying on these models. Anthropic announced in August 2025 that consumer data would be used for training by default unless users opted out, extending retention from 30 days to 5 years. OpenAI experienced significant outages due to a configuration error and an Azure datacenter power failure in December 2024, highlighting the operational risks of relying on third-party AI infrastructure. A Lock-in Scorecard evaluated five major LLM providers—OpenAI, Anthropic, Google Vertex, Mistral, and Cohere—across six dimensions affecting enterprise AI operations, including data policy, model stability, fine-tuning portability, API compatibility, operational risk, and contractual risk, with total scores indicating the level of lock-in risk. OpenAI scored lowest due to frequent model deprecations and outages, while Mistral scored highest due to its open-source model availability and data sovereignty focus, offering lower lock-in risk compared to other providers. The scorecard suggests that the choice of provider should depend on specific enterprise needs, such as data privacy, model stability, and migration costs, and recommends considering strategies like multi-provider routing or self-hosting infrastructure to mitigate lock-in risks.