Developer builds unified error model to tame inconsistent AI provider error responses

A developer building an API gateway that routes requests across OpenAI, Anthropic, and Gemini found that integrating the APIs themselves was straightforward, but handling their inconsistent error responses proved far more time-consuming. Each provider returns errors in different formats, and even identical HTTP status codes like 429 can carry different meanings depending on the provider. To address this, the developer replaced raw error forwarding with a normalization layer that maps all provider errors into a fixed set of internal categories such as rate_limited, overloaded, and quota_exceeded. This approach lets downstream logic decide whether to retry a request based on the failure type rather than which provider was called. The solution was eventually incorporated into Apiarium, an AI gateway the developer is building to abstract away provider-specific inconsistencies.