Production teams running AI at scale almost universally adopt multi-provider LLM routing by design. The core economic argument: a 60× price spread exists across model tiers on a single provider (from $0.25 to $15/M input tokens), and a 36× spread exists between a small model and Claude Sonnet for the same classification task. Routing traffic by workload type — batch to cheapest, real-time to lowest latency, compliance-sensitive to certified providers — is the rational response to this reality. The article covers three routing approaches: self-hosted LiteLLM, managed OpenRouter, and DigitalOcean's first-party Inference Router. Key tradeoffs discussed include the tension between routing and prompt caching (splitting a workload across providers can destroy a ~10× cache benefit for a ~2× rate saving), failover architecture design (secondary paths are cold by definition and may silently downgrade quality), and the 'goodput' metric as a better measure than raw throughput. OpenAI-compatible APIs across providers make switching a base URL change, not a multi-month migration.
