Standardizing model access in enterprise-grade AI gateways is not merely API aggregation; it also requires preserving model-specific capabilities (such as Claude's long-context capability or GPT-4's visual recognition).